Eve library

Manus 2.0 and Cue: competitor analysis and what Lilith should take (2026-09-30)

Research

Revision 19

On this page

Research document. Written 2026-09-30 by Claude (claude-mac-worker) at Elikem's request: "do a thorough competitor analysis and how we can fill any of the gaps in lilith".

  • How it was researched. A research agent read Manus's own pages first (the launch post, help centre, pricing page, Terms, Privacy and Usage policies, trust centre, API docs and cue.im), then the press. The author re-read the claims this document leans on; they are marked ✓ in section 11. Captures of the pages those claims rest on, made on 2026-09-30 and 2026-10-01, are attached in section 11.
  • Unverified claims. "(unverified)" means a claim appears only in a secondary or social source.
  • Reviewed. An independent read-only review (Codex, gpt-6.1-sol, xhigh) made 46 findings on the first draft. Every finding the sources support is folded in. Three rested on material the reviewer had not been given (board-only tasks and PRD §3); those facts are cited below to the pinned PRD or the attached board snapshot.
  • Status (updated 2026-10-01). As first written, this document changed no PRD and no board task, and the register in section 8 waited for Elikem. He approved it on 2026-09-30. A second independent review changed some details before filing; for example, P11 depends on P01 only. The comparison in sections 3 and 4 was made against V1.3 PRD revision 34. The filed changes are in V1.3 PRD revision 41 (approved; revision 34 is pinned and revision 41 is attached as an export in section 11), whose summary records that W12 (files view) and L02 (effort levels) were taken and the other optional items deferred or excluded under Decision 13. The new tasks A17, P11, R11, T12 and W12 were created on the board at 16:43–16:44 UTC on 2026-09-30 (board snapshot, section 11). Sections 7 to 9 are kept as proposed. Where they differ, PRD revision 41 and the board are authoritative.

0. Bottom line#

  • What launched. Manus 2.0 (2026-09-28) bundles:

    • a new harness, Cascade;
    • a paid always-on cloud computer;
    • event-triggered automations;
    • a desktop studio that edits video and builds multiplayer browser games;
    • remote control of your own desktop from your phone.

    Alongside it Manus launched Cue, a separate app of personal agents. Each agent has its own email address, phone number, wallet and computer, and several can work in one group chat. Cue is in free, invite-only early access. No independent hands-on review of 2.0 or Cue existed on 2026-09-30.

  • Timing is the largest gap. Cue arrived three weeks after Meta Muse (2026-09-08) and a day before OpenAI's dots (2026-09-29). Instinct, invite-only since August, raised $1B at a $10B valuation on 2026-09-28 (PRD §3). Before this analysis was filed, the lilith-v1-3 board held 213 items (193 tasks and 20 epics), all ready, and no task in the family has ever been claimed (board snapshot, section 11). The libraries the PRD builds on are listed in its §2.

  • The approved V1.3 plan already specifies most of what Cue offers:

    • an agent email address (A02);
    • a per-member workspace (W);
    • an agent wallet (P02);
    • outbound calls (A07);
    • standing duties (R08);
    • triggers (R02);
    • agent teams (X06).

    It also specifies controls Cue has not published: signed mandates, approvals bound to the exact action, principle 11, egress through Sentinel, a capability ledger and a flight recorder with proof.

  • Nine corrections to the approved plan (section 7A). These are needed whatever else is decided:

    • P02 must record that Link's agent wallet is open only to accounts created in the US and Canada.
    • Text written by other people must reach the planner only as quarantined data.
    • The sandbox must hold no reusable control-plane credential.
    • Agent mail needs rules on who may start work by email.
    • One provider may have several accounts, each with its own grants.
    • Automations need a real dry run and failure pauses.
    • The task estimate becomes a price ceiling, and allowance is restored after failures that were Lilith's fault.
    • Data terms and offboarding are published before the first invite.
    • Changes to approval defaults are versioned.
  • Three recommended additions (7B):

    • a number of her own that answers forwarded calls (A17);
    • a UK agent-payment route, found and assessed before any build (P11);
    • goal rooms in which several of Lilith's duties work on one goal with one writer per counterparty (R11).
  • Six optional items for Elikem to pick from (7C):

    • scan-to-task;
    • phone-to-desktop remote control;
    • code-rendered short video and share links;
    • effort levels;
    • a workspace files view;
    • a connector aggregator.
  • Not copied (7D, section 10):

    • selling cloud servers;
    • game hosting and an ad-video factory;
    • separate agent identities (Decisions 10 and 12 stand);
    • a public REST API, for now.

1. What Manus launched on 2026-09-28#

Manus announced 2.0 on X at 15:22 UTC on 2026-09-28 and introduced Cue five minutes later. It announced agent phone numbers at 17:43 UTC.

Component What it is Price and availability
Cascade Manus calls it its "in-house agent harness". It "keeps a project light from the start and brings in specialized capabilities only when the work needs them" ✓. The API docs mention Cascade "sub-lanes", whose connector confirmations are not exposed in the main task timeline, a cascadeAskUser wait state, and three agent profiles, standard, lite and max; a version in the profile name "is ignored because model versions are not selected independently". The claimed 23.2% fewer tokens, 28.2% less time and 32% lower cost ✓ "come from one tested configuration, not an overall average" (help centre). No test set, model or benchmark is published, and no Cascade release was found Included
Cloud Computer A persistent Ubuntu Server 24.04 VM with a dedicated public IP, SSH or web terminal, command line only, in a location the user picks. It first shipped on 2026-04-30. 2.0 presents it as the server behind multiplayer games and "a permanent home for an automation". If payment lapses, its files, tools and databases are deleted ✓ Standard $30/month (2 vCPU, 4 GB, 70 GB); Advanced $50/month (2 vCPU, 8 GB, 120 GB) ✓
Automations Three types: Schedule, Triggered task, and Advanced automation (beta), which turns a plain-language routine into a workflow the user reviews before it starts. Triggers: Gmail, Outlook, Mail Manus, Notion, Google and Outlook Calendar, GitHub, Shopify, Google Workspace, ad platforms, RSS and webhooks, with Slack named in the launch post. GitHub, Shopify, webhook and Mail Manus events can arrive as they happen; other sources are polled. There is a test run, conditions, and an automatic pause "after the same failure repeats across several checks" Runs use credits. Pro plans allow 20 scheduled tasks, Free allows 2
Manus Studio The desktop app (macOS and Windows) renamed. It handles documents, sheets, PDFs, slides, websites, code, games and video. Video Editor (desktop only) is a timeline for 30–60 second ads, AI UGC, motion graphics, tutorials and vlogs. Alchemy mode: "Manus takes on the creative direction, using video and code generation together". Game Dev makes browser games from templates, has an Edit panel, publishes as a playable link, and runs multiplayer on a purchased Cloud Computer ✓. There is also an AI music generator that Manus says makes "original, royalty-free music", exportable as WAV files. Users own their outputs. No generation models are named for 2.0 Credits. Manus's free-access campaign rules cap free users at 1 video and paid users at 10 videos a day, but the campaign's expiry date is ambiguous, so treat these as uncertain
Remote Control and Computer Use From the phone, by voice, Manus acts on your desktop in "its own visible, separate workspace", and the desktop streams live to the phone Desktop app
Cue A standalone app. "Each agent gets its own email, phone number, wallet, and computer" ✓. It "can also take your calls and leave you a summary". "Put several agents in a group chat with a shared goal, and they hand work to each other" ✓. "At a restaurant, scan the QR code and your agent can order for you or hold your place" ✓. A tile reads "Scan & create: Scan to create an agent." The connectors shown are Trip.com, Oura, Gmail and MyFitnessPal, plus "More apps"; the example prompts also use Plaid, Slack, Taskrabbit and Link. The FAQ says: "You control what it can access and which actions need approval." Free during early access with invite code MEETCUE ✓ (secondary reports say the code was capped at 1,000 users; unverified). cue.im offers macOS and Windows downloads and a mobile app; RuntimeWire reports web, desktop and mobile access, with iOS pending App Store review. No price after early access was found
Agent phone numbers "Manus & @CueAgents can now have their own phone number so your agent can make and receive calls and texts. You can even forward your calls to your agent! Available in select countries for now, with voice-only support in some." No country list was found in the sources searched (as of 2026-09-30) Not priced
Agent wallet Manus names no issuer. Cue's example prompts ("Pay with Link, and I'll confirm on my phone"; "Charge it to my Link") suggest Stripe Link; the issuer and integration are unconfirmed. Link's agent wallet itself ✓: "only available to users who created their account in the US and Canada"; the member approves each purchase biometrically within 10 minutes; the agent then gets a one-time card; the daily limit "can't be increased" See the Link FAQ

What 2.0 builds on (Manus 1.x, 2025–2026):

  • Wide Research, in which each sub-agent runs a full Manus with its own context.
  • Full-stack web apps with hosting, custom domains and Stripe.
  • Agents in Telegram and Slack. Secondary sources also mention WhatsApp Business and LINE; no Manus page confirming them was captured (unverified).
  • Mail Manus: a personal inbox, approved senders, and workflow addresses with preset instructions.
  • Several Gmail and Google Calendar accounts per user.
  • The Browser Operator extension.
  • Projects that learn.
  • Agent Skills.
  • Plan Mode, in which a plan is confirmed before execution.
  • Connectors to ElevenLabs, Higgsfield (MCP), Zoom, Shopify and Supabase.
  • Zapier ("8,000+ apps").
  • A REST API v2 with agents, tasks, webhooks and OAuth2.

These come from Manus's blog index, help centre and docs (captures in section 11).

Plans. The published prices are in USD, and a help article says every price is set in USD. Stripe's case study says Manus charges in 31 currencies, including GBP; the sources conflict.

Plan Price What you get
Free $0 300 daily credits, usable on 2.0 Lite only; daily credits capped at 1,500 a month; 1 concurrent task; 2 scheduled tasks
Pro $20/month 4,000 credits a month
Pro $40/month 8,000 credits a month, with a menu of larger allowances
Pro $200/month 40,000 credits a month, with a menu of larger allowances
Team US$20 per seat a month 4,000 credits per seat in a shared team pool; SSO available
Business use Master Services Agreement The Terms say it governs the Team plan, the APIs and other business services
  • The three Pro prices are from the pricing page as rendered on 2026-10-01 (screenshot in section 11); the prices behind the larger-allowance menus were not captured.
  • All Pro tiers allow 20 concurrent and 20 scheduled tasks. Annual billing saves 17%.
  • Cloud Computer is billed on top.
  • Credits are refunded only for tasks that fail "due to technical issues on our end" (credits article). Refund requests "are carefully reviewed by a specialized AI agent and our Customer Success team" (refund-policy article).
  • For Wide Research, Manus shows an estimate that must be confirmed first, and "will absorb the additional cost" if the run exceeds it (trust centre).

2. The company#

Manus is Butterfly Effect Pte. Ltd. of Singapore. Its privacy policy names Butterfly Effect as controller under European data protection law and EDPO as its EU and UK representative. Its trust centre says data is stored in the US and Singapore.

Timeline:

Date Event
2025-12-29 Meta announced the acquisition (about $2bn)
Early 2026 Chinese regulators opened an investigation into the deal
Mar 2026 The co-founders were reportedly barred from leaving China
2026-04-27 (BBC) The National Development and Reform Commission ordered the deal withdrawn
2026-08-11 Manus announced a return to independence
2026-08-23/24 Manus deleted data created after 2025-12-29 for users in "affected regions", which it does not name
2026-09-01 Manus formally resumed independent operations under its founding team

Ownership now (all secondary, unverified):

  • The founders, Tencent, HSG and ZhenFund bought Meta out at about $2bn.
  • Tencent is reported to be the largest external investor.
  • A $500m raise at a $4bn valuation is in progress.
  • A Hong Kong IPO restructuring is being considered.
  • A mainland-China product is being prepared.

Revenue: Manus's own figures were $100M ARR and a $125M run-rate in December 2025. The FT reported "close to $500m" in July 2026 (unverified).

Why it matters to Lilith. Cue asks members to hand over a phone line, an inbox, bank reads (Plaid) and a card. The privacy policy dated 2026-09-28, in its US opt-out rights, still describes "sharing identifiers about you with our marketing vendors to promote Meta products and services on other platforms". Lilith is a UK company with a UK data controller. It can publish where each class of member data lives and who processes it, a question on which Manus's answers are still moving (7A-8).

3. Cue against Lilith V1.3, feature by feature#

Sources for the Lilith column: the V1.3 PRD as approved at revision 34 (Eve 8824516e, pinned), which this comparison was made against. X10 and K11 exist only on the lilith-v1-3 board (board snapshot, section 11). Revision 41 later changed some rows; see the status note. Nothing in that column is built yet.

How to read the verdicts:

  • "specified" means the Lilith plan states the control.
  • "unpublished" means Manus's sources found no statement either way.
Capability Cue / Manus (2026-09-30) Lilith V1.3 plan Verdict
Agent email address Each agent has its own address; the domain is unpublished. Mail Manus has approved senders and workflow addresses A02 (M1): <handle>@agent.oshunsynthetics.com, persona name with Lilith behind it (Decision 10). Who may start work by mail is not specified Parity on the address; correction needed on who may start work by mail (7A-4)
Phone, outbound Its own number; calls and texts in unnamed "select countries" A07 (M2): outbound calls with disclosure, IVR, hold, take-over, transcript; Decision 4 (provider, countries) open Parity in plan, later in time
Phone, inbound and forwarded calls "Take your calls and leave you a summary"; forward your own calls to the agent Nothing. S03 (M2) lets the member text Lilith; it does not answer callers Gap (7B-1)
Wallet Cue's examples suggest Stripe Link; issuer unconfirmed P02 (M2) Stripe Link agent wallet; P03 UCP + AP2; P04 merchant hand-off by default; P09 x402 (M3); Decision 5 open Same candidate provider; P02 does not record Link's account-origin limit (7A-1); no UK-eligible route has been assessed (7B-2)
Spending controls "Pay within the budget you set"; member-set approval boundaries; RuntimeWire notes the announcement does not explain transaction approval or after-the-fact review P01 signed mandates, P06 guardian mandates, P07 regret window, T09 approvals bound to the exact action, P05 receipts, P10 provider-backed protection Lilith specified; Cue's detail unpublished
Agent computer "A dedicated workspace" (spec unpublished; the Manus Cloud Computer is Ubuntu CLI) W01–W05: per-member microVM, Sentinel egress gate, surrogate-credential vault, taint tracking Lilith specified; Cue's isolation unpublished
Several agents on one goal A group chat of agents with a shared goal, handing work to each other; the member makes the final call X06 agent teams with per-teammate budgets; X10 sub-agent result check; R05 plans; R08 standing duties with their own grants; S11 side chats; W07 activity. There is no combined view of one goal, no handoff semantics and no coordination rule for outbound writes Gap in the view and in coordination (7B-3)
Agents per responsibility "Create distinct agents for different responsibilities" R08 standing duties named by the member, each with its own grants, mandate, budget and triggers; one Lilith (Decision 12) Same job, different model; Decision 12 stands
Connectors Trip.com, Oura, Gmail, MyFitnessPal and "More apps" shown; example prompts use Plaid, Slack and Taskrabbit; Manus supports several Gmail and Google Calendar accounts C03 Google, C04 Microsoft, C06 health (Oura), C07 lifestyle and booking (Instacart named), C08 open banking, C09 MCP client. Trip.com, Taskrabbit and MyFitnessPal are not named. How many accounts one provider may have is not specified Mostly parity; correction on several accounts per provider (7A-5); the rest go to C07's list when a member cohort needs them
Routines "Connect the services ... then define a routine" R02 triggers, R05 plans, R06 ritual briefs, R08 duties Parity in plan; R02 lacks dry-run and failure semantics (7A-6)
QR in the physical world Scan a restaurant QR code to order or hold a place Nothing (I13 uses QR only for device pairing) Optional (7C-1)
Approvals "You set the review boundaries" Principle 3, T09, T11 caution levels, principle 11, X03 action monitor, W07 flight recorder with proof Lilith specified; Cue's detail unpublished
AI disclosure The Terms (§2.13) say Manus "may include technical or legally required identifiers and disclosures" and make users responsible for consents and required notices ✓; the Usage Policy forbids unlawful impersonation and missing required disclosures. No universal commitment that agents say they are AI was found Principle 9 and Decision 10: disclosure on every outbound email, message and call. Inbound calls are not yet covered (7B-1) Lilith commits; Manus defers to the law
Memory Unpublished on cue.im M01–M07: flow labels, weekly review, correction propagation Lilith specified; Cue's design unpublished
Platforms macOS, Windows and mobile (cue.im); iOS pending App Store review (RuntimeWire) Web now; iOS/Android (V1.1); desktop, Linux/Omarchy, de-Googled Android, CLI, self-hosted node and extension (I, M1–M3) Broader in plan, later in time
Price Free during early access; later price unknown Q01 compute tiers; Decision 6 open; conversation not metered Unknown on both sides
Age and content Manus is "not intended for use by anyone under 18" (Privacy Policy); the Usage Policy bans adult content, "such as erotic chat, pornography" ✓ 18+ (M1 personal invites); mature audience with hard lines Different stances; matters more for storytelling features than for the agent

4. The rest of Manus 2.0 against Lilith#

Capability Manus 2.0 Lilith Verdict
Loading capabilities on demand (Cascade) Design unpublished; one efficiency figure with no configuration X06 tool search, code mode and agent teams, measured on E01; K01 progressive skill loading; K11 skill retrieval picking at most two (board, added 2026-09-29); L02/L03 routing and a verifier-gated cascade Covered. Lilith's done-whens publish measured deltas
Agent tiers Lite / standard / max; Free gets Lite only L02 cheapest-capable router; Q01 tiers Optional effort levels (7C-4)
Cost ceiling Wide Research: estimate confirmed first, overrun absorbed by Manus L06 estimate and receipt; L07 budgets Correction: the estimate becomes a ceiling (7A-7)
Persistent cloud computer $30–$50/month VM with public IP and SSH, hosting bots, sites and game servers W02 per-member workspace with idle suspend; R01 durable background tasks Different on purpose; no hosting (7D). Optional files view (7C-5)
Event automations Pushed and polled triggers; test run; pause after repeated failure R02 triggers with fixtures Correction (7A-6)
Remote control from the phone Voice from the phone; desktop acts and streams live S10 desktop local actions (M2) behind the W10 local Sentinel; I13 pairing; A04 watch-live for the cloud browser Optional (7C-2)
Documents, sheets, slides, web apps Yes, with hosting, custom domains and Stripe A15 (M2): documents, sheets, decks and small apps; no share link Optional share links (7C-3); no hosted apps
Video editing and code-rendered video Video Editor timeline; Alchemy (code plus video) Members: nothing. V1.5's "code-as-shot" (Y06) serves Elikem's studio (V1.5 PRD b40979b3 r36) Optional (7C-3)
Browser games, multiplayer Game Dev on purchased servers V1.4 Z03 games channel (M4); V1.5 Z06 game tie-ins (M4) Not in V1.3 (7D)
Messaging channels Telegram and Slack as agent channels; WhatsApp Business and LINE reported but unverified S01 Telegram (M1); S02 WhatsApp once its terms are verified; S03 SMS (M2). Slack is a connector, not an agent channel; no LINE Telegram parity; WhatsApp conditional; Slack and LINE channels deferred until a cohort needs them (7D)
Developer surface REST API v2 (tasks, agents, webhooks, OAuth2), Zapier C09 MCP client (M1); I11 CLI with MCP server mode (M1); D01–D03 connector platform and A2A (M3). No public REST task/agent API or inbound webhooks Public API deferred (7D); aggregator optional (7C-6)
Compliance SOC 2 Type 2 attestation, ISO/IEC 27001:2022 and ISO/IEC 27701:2019 (trust centre); nine named sub-processors; data in the US and Singapore No certification commitment; published data terms not yet planned Correction: publish data terms (7A-8). Certification is a separate later choice (section 9)

5. Where Manus is weak, and where it is not#

  1. No universal AI-disclosure commitment. The Terms defer to "legally required" disclosures and put consent duties on the user ✓. Lilith's principle 9 commits to disclosure on every outbound contact. Inbound calls are not yet covered (7B-1).
  2. Two reported injection vulnerabilities, both patched. These were researcher demonstrations; no in-the-wild exploitation was reported.
    • Salt Labs (via Dark Reading, OODA Loop and Implicator; the primary text was not retrieved): one email carrying obfuscated JavaScript got past a plaintext filter, ran in the victim's Manus environment and pulled connected-account tokens. It was disclosed 2026-09-24 and patched through Meta's bug bounty.
    • CodeAnt AI (its own write-up, not independently verified): a shared Project's instructions loaded as trusted configuration into other members' agents, giving code execution in their sandboxes and a remote-desktop takeover. The remote-desktop admin password sat in the sandbox environment. Fixed 2026-09-16.
    • Gartner, quoted in InfoWorld on Manus's remote execution and system access: "the autonomy is ahead of the governance". The point is about enterprise governance, not about Cue specifically.
  3. Its approval documentation is inconsistent. The desktop launch post (2026-03-16) says "Every terminal command requires your explicit approval before execution", and that users can choose "Always Allow" for trusted tasks. A help article dated 2026-03-24 says that "once a user authorizes a folder, Manus runs commands automatically" and asks only for sensitive ones. The two pages differ in wording. They do not show whether the default behaviour changed, when, or whether users were told.
  4. Training and residency.
    • The privacy policy allows aggregated, de-identified or anonymised data to be used "to develop and improve our models".
    • The Team page says Manus "explicitly prohibits model providers from training on" Team and Enterprise customers' data.
    • The trust centre says formal agreements with all third-party model providers ensure "they do not use your data for training their models".
    • Data is stored in the US and Singapore.
  5. Credits. Manus's credit refunds cover platform faults only, and monthly credits do not roll over. The recurring complaints are credit burn, credits lost on downgrade and slow human support (Trustpilot and Reddit; unverified).
  6. Where Manus is strong and Lilith should match it:
    • it confirms the estimate and absorbs overruns for Wide Research;
    • UK and EU consumers get a 14-day withdrawal right and the courts of the country where they live (Terms);
    • it holds independent certifications and publishes a named sub-processor list.
  7. The UK is thinly served by identity-bearing agents:
    • Link's agent wallet is open only to accounts created in the US and Canada ✓;
    • Cue's numbers work in unnamed "select countries";
    • Muse is reported to be US-only, and dots is reported to be closed to UK, EEA and Swiss personal Pro plans. Both come from PRD §3, which cites secondary sources for them because OpenAI's own pages could not be read on 2026-09-29; neither is verified here.
  8. Parallel writes. Cue lets several agents, each with email, a phone line and a wallet, work on one goal. Whether Cue coordinates their outbound actions is unpublished. For any design like this, two writers can contact the same person or charge the same merchant, and those effects cannot be recalled (7B-3). An opinion piece (vocal.media, unverified) argues this conflicts with Manus's earlier stance; that claim is not relied on here.

6. What the plan specifies, and the timing problem#

The approved plan specifies controls that Manus has either not published or handles differently:

  • containment: Sentinel, surrogate credentials and taint tracking;
  • honesty records: a capability ledger reconciled against live scopes, and a flight recorder with proof;
  • spending: mandates, guardian limits and a regret window;
  • principle 11: unasked work only reads;
  • outbound disclosure;
  • wellbeing policy;
  • model sovereignty;
  • reach: Linux, Omarchy, de-Googled Android and the self-hosted node.

These are specifications, not measurements. Neither side's controls have been tested against the other's.

The timing problem.

  • No V1.3 task has started. Muse, Cue and dots ship today, and Instinct is inviting users.
  • What makes Cue feel like a person is an address, a number, a wallet and work that continues while you're away. In Lilith these land at M1 (email, triggers, duties) and M2 (calls, SMS, wallet).
  • The M1 cohort is Elikem's personal invitees, so the pressure is to demonstrate the approved plan well, not to win share. That is why section 7 separates corrections, which are cheap and make M1 safer, from new scope.
  • Decision 9 (sequencing) already allows V1.3 M0 and M1 to run in parallel with V1.1 and V1.2. This analysis supports doing so.

7. What to change#

These change task bodies already on the board. None adds a milestone.

7A-1. P02: record Link's eligibility rule.

  • Link's agent wallet "is only available to users who created their account in the US and Canada" (support.link.com, read 2026-09-30 ✓). The rule is where the Link account was created, not where the member lives.
  • P02 reads eligibility from Link's own OAuth or account response. It does not infer eligibility from the member's location.
  • An ineligible account fails closed with the provider's reason and falls back to P04 merchant hand-off.
  • Done when: a test-mode ineligible account is refused with the reason shown, and P04 is offered.

7A-2. T01 and F03: text written by other people is data, never instructions.

  • Content another person wrote reaches the planner only through the quarantined parser. This covers mail, shared lists, plans and rituals, household members' notes, calendar invite text, group messages and goal-room messages from others.

  • Flow labels are kept across handoffs.

  • Authenticated requests from the member stay distinct from third-party content.

  • T01 (M0) tests this with fixtures:

    • an instruction planted in a shared-list fixture;
    • obfuscated and encoded payloads in mail (JSFuck-style JavaScript, base64, homoglyphs).

    Done when: the injected content cannot change trusted instructions, gain a permission, run its payload or cause an unauthorised effect, while the legitimate task still completes.

  • F03 (M3) adds the same case against the real shared-list surface and gains a dependency on T01.

7A-3. A04 and W02: the sandbox holds no reusable control-plane credential.

  • The remote-browser stream server's admin credential, and every control-plane secret, live outside the member sandbox.
  • Done when a hostile process inside the sandbox cannot obtain a reusable control-plane credential or call a privileged operation. The test searches environment variables, readable files, mounts, process metadata and the control-plane endpoints the sandbox can reach.

7A-4. A02: who may start work by mail.

  • The member chooses approved senders for the agent address.
  • Mail from anyone else is third-party content under 7A-2. It never creates a grant and never becomes a member request.
  • A standing duty may use a sub-address, so replies thread back to that duty. The disclosure and persona signature are unchanged.
  • Done when these cases are handled correctly: a spoofed approved sender (failed SPF/DKIM/DMARC), a forwarded mail carrying instructions, and a CC'd third party.

7A-5. C02: several accounts per provider.

  • A member may connect several accounts of one provider, for example two Gmail accounts. Manus supports this.
  • Each grant is bound to one account, and every send names the account it uses.
  • Done when a send from the wrong account is refused.

7A-6. R02: automations the member can trust before switching them on.

  • Dry run. It runs through the tool gate with every write simulated: zero live effects, including reservations and read-status changes such as marking mail read. It shows what would have been sent, booked or paid.
  • Failure handling. Failures are classified (transient, auth, provider-side, own bug). After the same class repeats a set number of times, the trigger pauses, with backoff and a message naming the failure. The threshold, reset and notification are part of the spec.
  • Speed. Each trigger states whether it fires on pushed events or on polling, and the interval.
  • Activation card. A plain-language routine becomes a card listing trigger, conditions, actions, accounts, grants and budget. Activating it signs only the mandate shown on the card. Anything outside the mandate still waits for T09 approval.
  • Done when:
    • a dry run of a send-and-book routine produces zero external effects;
    • a repeated auth failure pauses the trigger;
    • an action outside the card's mandate waits for approval.

7A-7. L06, L07 and Q01: fair billing.

  • The estimate is a ceiling. The estimate shown before a task (L06) is the most it may charge. Going over needs fresh approval. An overrun caused by Lilith's own estimating error is absorbed; Manus already does this for Wide Research.
  • Allowance is restored after Lilith's own failures. After the external effects are reconciled (A13), a failure is given a deterministic reason code. When the code says Lilith was at fault (platform error, own bug), the task's compute returns to the member's allowance, shown on the receipt, with no claim to file.
    • This restores Oshun-metered allowance only. It is not a cash refund, and it cannot touch a member's own provider bill (BYOK).
  • No silent removal. A paid allowance is never removed silently on a plan change.
  • Currency. UK members see prices in pounds.
  • Budget reservation needs proof. capacity-budget.ts reserves before dispatch today. Before the terms promise that a balance can never go negative, tests must cover concurrent dispatch, retries and reconciliation of the actual cost.

7A-8. New task T12 "Published agent data terms and offboarding" (M1, before the first invite).

A versioned, plain-language page, reviewed by counsel. For each class of member data (workspace, memory, flight recorder, mail, calls, payments) it states:

  • where the data is stored and processed;
  • where it is transferred;
  • the named processors and admitted model routes;
  • the training posture: Oshun-paid routes never train on member data (§6.12), and member-owned routes show theirs;
  • the purpose and retention of third parties' data (correspondents, and callers once 7B-1 ships).

It also covers:

  • Export and erasure, as T06 implements them.
  • Portability of operating state. It says what export restores (plans, duties, artifacts, provenance) and that grants need fresh authorisation after a restore. IDC's point, quoted in InfoWorld: "an export button alone does not establish that an agent's operating state is portable".
  • Offboarding. It states the notice and grace period before a workspace or its files are deleted. Nothing is deleted abruptly, unlike Manus's Cloud Computer, which deletes files when payment lapses.
  • Consumer terms. Oshun's own subscription, cancellation and withdrawal terms for UK and EU consumers, which compute allowances and purchase protection do not replace.

Depends on G01 and T06. Its done-when: the page exists, is linked from signup and settings, every statement is checked against the code and the processor list, and the version is recorded.

7A-9. T11: approval defaults change only by versioned decision.

  • A change to a caution level's or an approval class's default is versioned.
  • Members are told in the app.
  • It never applies to existing grants or mandates without fresh consent.
  • Done when changing a default leaves existing grants unchanged until the member consents.

7B-1. New task A17 "Her own number: forwarded calls answered, screened and summarised; texts from her number" (M2, effort L).

What Cue and Manus agents do: they have their own numbers and take forwarded calls. Instinct is a texting-first agent with its own number (PRD §3).

Who reaches her

  • The member sets up a duty that answers calls: forwarded calls from their own number (on busy or no answer), or calls to her number. Automatic answering runs only under a signed, bounded mandate. The mandate lists what she may tell a caller and which callers she may tell it to. Flow labels (M01) still apply.

What she says and does on a call

  • Every call opens with the disclosure and the persona, for example "You've reached Morgan, Sam's AI assistant from Lilith". This extends principle 9 to inbound calls.
  • She takes messages and answers only within the mandate.
  • A commitment (a booking, a payment, a promise) needs an exact-action approval or a covering mandate.
  • What a caller says is untrusted content (7A-2). It cannot start further work.
  • Each call ends as a summary in the flight recorder (W07). The member's real number is never disclosed without consent.
  • Spam and robocall screening, plus a block list.

Legal and regulatory

  • Counsel approves the purposes, notices, recording and retention rules for each supported jurisdiction. Where those conditions cannot be established, the default is no recording.
  • Number-registration duties (for example US carrier registration for application texting) are confirmed with the chosen provider, with dated evidence.
  • Texts from her number follow A11's approval rules.

Tasks and done-when

  • Depends on A07, S03, T04, M01, W07 and X09; outside, Decision 4.
  • Done when, on test numbers:
    • a forwarded call is answered with the disclosure;
    • a caller's request for information outside the mandate is declined;
    • a caller's instruction to the agent does nothing;
    • the summary lands in the flight recorder.

7B-2. New task P11 "Agent payments for members outside the US and Canada: assess and recommend" (M2, effort M, discovery only).

The assessment covers each candidate route:

Candidate Status in this assessment
Link If and when it admits accounts created elsewhere
Card-network agent programmes (Visa Intelligent Commerce, Mastercard Agent Pay) Through a PSP that offers them in the UK
Open-banking payment initiation Through a regulated provider, approved in the member's bank app; TrueLayer is already the C08 candidate for read-only banking
The V1.0 crypto rail (P09) and merchant recognition (P08) Evaluated only here, since both are M3

For each route it records:

  • dated evidence of UK availability;
  • Oshun's regulatory role, for example whether any route would make Oshun a payment or e-money business;
  • the purchase protection a provider actually backs, as principle 4 requires;
  • how the route connects to receipts (P05), disputes (P10) and the regret window (P07).

The output is a recommendation for Decision 5. A build task is filed only after Elikem decides, and P04 merchant hand-off stays the default meanwhile.

Depends on P01, P04 and P10; outside, Decision 5.

7B-3. New task R11 "Goal rooms" (M2, effort M).

It builds on the existing machinery: R05 plans, R08 duties, X06 agent teams, X10 result checks, W07 activity and S11 side chats. It adds three things.

  • One view per goal. Every duty and teammate working on a goal appears in one thread as a lane the member can name ("Bills", "Trip"). Each lane is still Lilith and signs as the persona, within Decisions 10 and 12. The member sets direction and makes the final call there.
  • Handoff semantics. A handoff is a typed record: from, to, what, why. It passes work, never grants. The receiving lane acts only within its own grants (R08).
  • One writer per counterparty.
    • Outbound actions carry a member-scoped conflict key: recipient address or number, merchant, calendar.
    • A second lane's action on the same key waits, visibly, and re-checks its grants and approvals once it may proceed.
    • Keys and waits are part of A13's idempotency and reconciliation, so a timeout, crash or lease expiry cannot produce a second external effect.
    • One member's keys never block another member.
  • Depends on R05, R08, X06, X10, W07, S11 and A13.
  • Done when:
    • two lanes emailing the same person produce one send and one visible wait;
    • a crash after commit recovers to exactly one effect;
    • a handoff does not widen the receiver's grants.

7C. Optional scope (Elikem picks; none is needed for M1)#

  • 7C-1. S13 "Scan to hand over" (M2, effort S).
    • The native app's camera reads a QR code, menu, letter or form into a task.
    • Only https links are opened, and each goes through Sentinel's redirect-chain and private-network checks.
    • A page that asks for credentials is refused.
    • Nothing is kept beyond the task without consent.
    • Joining a queue or ordering is a write, so it needs approval or a mandate.
    • Done when these cases are handled: a QR code carrying instructions, a private-network destination, a deceptive redirect, and misread text (OCR errors surfaced for the member to confirm).
    • Depends on S04, A04, T01 and W03.
  • 7C-2. S10 amendment: phone-to-desktop tasking.
    • From a paired phone, the member asks Lilith to act on their paired desktop and watches live.
    • Each local action still passes W10 and needs its approval.
    • During take-over, capture, streaming and model access are suspended; they resume only under control.
    • Pairing, revocation and session authentication come from I13.
    • Done when a take-over produces no captured frames and no trace data holding secrets.
    • Adds dependencies on I13 and S04.
  • 7C-3. A15 amendment: short videos rendered from code, and share links.
    • A15 gets its own deterministic renderer in M2 for explainers, motion graphics and captioned clips, in vertical or wide. No generative video model is needed, and outputs are marked under T10. V1.5's Y06 may later share it; A15 does not wait for Y06.
    • Share links are static, view-only previews:
      • approval is bound to the artifact version, audience and expiry;
      • flow labels are enforced and secrets stripped;
      • a link can be revoked, and cached copies must honour that.
    • Hosted, executable apps are out.
  • 7C-4. L02 amendment: effort levels.
    • The member may choose quick, standard or thorough.
    • Each level is a measured quality, latency and cost profile above the mandatory floor.
    • Every level still requires E07 intake, grants and budgets, and fails honestly when no admitted route fits.
  • 7C-5. New task W12 "Workspace files view" (M1, effort S).
    • The member sees the files Lilith saved in their workspace, and can export or delete them.
    • Paths are confined to the member (symlink-safe).
    • Previews are inert, and flow labels apply.
    • Deleting a file never corrupts evidence the flight recorder holds; it reuses T06's erasure path.
    • Done when a cross-member path, a hostile filename and hostile HTML content are all handled safely.
    • Depends on W02, W06, W07 and T06.
  • 7C-6. C09 amendment: evaluate one MCP aggregator (for example Zapier's).
    • Admitted only if Sentinel can see each downstream service, action, scope and data flow.
    • Tool definitions are pinned; revocation and auditable outcomes are required.
    • Opaque routing is rejected.
    • Done when a changed tool definition is caught and an unauthorised downstream call is refused.

7D. Deliberately not taken (for now)#

  • Selling cloud servers.
    • A $30–$50 VM with a public IP is a hosting business and an abuse surface.
    • R01 durable tasks carry standing duties without a VM that is always on.
  • Game Dev, multiplayer servers and an ad or UGC video factory.
    • These are far from the wellbeing-first agent.
    • Games stay on the V1.4 and V1.5 horizons (Z03, Z06).
  • Separate agent identities per responsibility.
    • Decision 12 (still Lilith underneath) and Decision 10 (outbound identity) stand.
    • Goal-room lanes (7B-3) cover the job.
  • A public REST task/agent API with inbound webhooks, and Slack or LINE as agent channels.
    • Deferred until a cohort or a D-pillar partner needs them.
    • No task is proposed now.
  • Members in the V1.5 studio before its M4 horizon. 7C-3 covers the everyday need.

8. Change register (proposed, not yet made)#

# Type Key Change Milestone Depends on Decision needed
7A-1 Correction P02 Link account-origin eligibility; fail closed; fall back to P04 M2 — —
7A-2 Correction T01, F03 Other people's text is quarantined data; fixtures in T01, real surface in F03; F03 depends on T01 M0, M3 — —
7A-3 Correction A04, W02 No reusable control-plane credential reachable from the sandbox M1, M0 — —
7A-4 Correction A02 Approved senders; third-party mail stays data; duty sub-addresses M1 — —
7A-5 Correction C02 Several accounts per provider; account-bound grants M0 — —
7A-6 Correction R02 Dry run at the tool gate; failure classes and pause; speed label; activation signs only the shown mandate M1 — —
7A-7 Correction L06, L07, Q01 Estimate is a ceiling; allowance restored on Lilith's own failures; no silent removal; GBP; prove reservation M1, M3 — Decision 6 (prices)
7A-8 New task T12 Published agent data terms and offboarding M1 G01, T06 —
7A-9 Correction T11 Approval-default changes versioned, announced, re-consented M1 — —
7B-1 New task A17 Her own number: forwarded calls answered, screened and summarised; texts M2 A07, S03, T04, M01, W07, X09 Decision 4
7B-2 New task P11 Agent payments outside the US and Canada: assess and recommend M2 P01, P04, P10 Decision 5
7B-3 New task R11 Goal rooms: one view, handoff semantics, one writer per counterparty M2 R05, R08, X06, X10, W07, S11, A13 —
7C-1 Optional task S13 Scan to hand over M2 S04, A04, T01, W03 Decision 13
7C-2 Optional amendment S10 Phone-to-desktop tasking; capture suspended during take-over M2 + I13, S04 Decision 13
7C-3 Optional amendment A15 Code-rendered short video; static share links M2 — Decision 13
7C-4 Optional amendment L02 Effort levels above the floor M1 — Decision 13
7C-5 Optional task W12 Workspace files view M1 W02, W06, W07, T06 Decision 13
7C-6 Optional amendment C09 Evaluate one MCP aggregator M1 — Decision 13

PRD text:

  • §3 gains a Manus 2.0 / Cue row. Comet and Genspark keep their rows.
  • Decisions 1, 4, 5 and 9 gain the inputs below, and Decision 13 is added.
  • §6 sections gain the corrected paragraphs.
  • Each new task gets its route-matrix and walkthrough rows where it adds a page, and its epic's milestone exit check.
  • The epics A, P, R and T gain their new tasks; S and W do so if 7C is taken.

9. Decisions this raises for Elikem#

  1. Decision 1 (regions), input. The phone and payment features of these agents are centred on the US and Canada:

    • Link's agent wallet admits only accounts created there ✓;
    • Cue's numbers work in unnamed countries;
    • Muse is reported to be US-only, and dots reported to exclude UK personal Pro plans (PRD §3, secondary sources; unverified).

    The recommendation below rests on the verified Link rule and on Lilith's own plan, not on the reported Muse and dots limits.

    Recommendation: admit UK members for M1's capabilities (mail, calendar, browser, duties) once T12 and the M1 exit pass. UK numbers (A17) and a UK payment route (after P11) follow at M2, which also needs V1.1.

  2. Decision 4 (telephony), input. The provider must support inbound and forwarded calls and texts from the agent's number, not only outbound calls.

  3. Decision 5 (commerce order), input. Link admits only accounts created in the US and Canada. If Link goes first, those members go first. P11's recommendation informs the choice for everyone else.

  4. Decision 9 (sequencing), input. The category shipped in September. Recommendation: run V1.3 M0 and M1 in parallel with V1.1 and V1.2, as the PRD allows.

  5. New Decision 13: which optional items (7C-1 to 7C-6) to take into V1.3 now. Recommendation:

    • take 7C-5 (files view, cheap and trust-building) and 7C-4 (effort levels);
    • hold 7C-1, 7C-2, 7C-3 and 7C-6 until the M1 cohort shows demand.
  6. Later: independent certification (SOC 2, ISO 27001/27701) before outside cohorts beyond personal invites. Manus holds these, and T12 does not substitute for them.

10. What not to copy#

  • Several agents that each hold a phone line, inbox and wallet with no coordination rule. Emails, calls and payments cannot be recalled (7B-3).
  • Disclosure left to what the law requires. Lilith discloses every time, inbound too once A17 ships.
  • Approval behaviour documented inconsistently, with no dated record of a change. Lilith versions its defaults and asks again (7A-9).
  • Shared text treated as trusted configuration. Other people's text is data (7A-2).
  • Secrets in the sandbox environment (7A-3).
  • Efficiency figures from "one tested configuration". Lilith publishes the task set with any claim (E, X06).
  • Selling hosting; game and ad factories (7D).
  • Workspace files deleted when payment lapses (7A-8), and refund rules that cover platform faults only (7A-7).

11. Sources (read 2026-09-30)#

✓ = re-read by the author on 2026-09-30.

Captures attached (Eve media; each records its URL, final URL, HTTP status and capture time):

Manus: product and help centre

  • ✓ Launch post: https://manus.im/blog/introducing-manus-2-0 (Cue identity, group chat, QR, Cascade figures, MEETCUE, multiplayer steps)
  • Help centre, "What is new in Manus 2.0?": https://help.manus.im/en/articles/17190150-what-is-new-in-manus-2-0 ("one tested configuration, not an overall average"; Cue access)
  • ✓ Cloud Computer plans: https://help.manus.im/en/articles/15392078-understanding-cloud-computer-plans-and-billing ($30/$50; Ubuntu Server 24.04; deletion on lapse)
  • Cloud Computer launch (2026-04-30): https://manus.im/blog/manus-cloud-computer
  • Pricing: https://manus.im/pricing (screenshot attached)
  • Team plan: https://manus.im/team
  • Membership pricing: https://help.manus.im/en/articles/11711111-what-is-the-current-membership-pricing-for-manus
  • Credits: https://help.manus.im/en/articles/11711097-what-are-the-rules-for-credits-consumption-and-how-can-i-obtain-them
  • Credit refund policy: https://help.manus.im/en/articles/12992237-how-does-our-ai-agent-s-credit-refund-policy-work
  • Currency: https://help.manus.im/en/articles/11711125
  • Free-access campaign rules: https://help.manus.im/en/articles/16312548-manus-free-access-campaign-rules
  • Automations: https://manus.im/docs/automations.md
  • API v2: https://open.manus.ai/docs/v2/task-lifecycle.md (sub-lanes, cascadeAskUser), https://open.manus.ai/docs/v2/task.create.md (agent profiles, model versions)
  • Zapier: https://manus.im/docs/integrations/zapier.md
  • Blog index (connector and feature posts): https://manus.im/blog
  • AI music generator: https://manus.im/tools/ai-music-generator
  • Mail Manus: https://help.manus.im/en/articles/12059745-how-can-i-use-mail-manus
  • Desktop launch (2026-03-16): https://manus.im/blog/manus-my-computer-desktop
  • Desktop auto-run help article (2026-03-24): https://help.manus.im/en/articles/14178443

Manus: policies and trust

  • ✓ Terms (dated 2026-09-28; §2.13 disclosures): https://manus.im/terms
  • Privacy policy (dated 2026-09-28): https://manus.im/privacy
  • ✓ Usage policy (dated 2026-09-28; adult content banned): https://manus.im/usage
  • Trust centre (certifications, sub-processors, data location, Wide Research overruns): https://trust.manus.im/, https://trust.manus.im/faq and https://trust.manus.im/subprocessors

Cue and announcements

  • Cue site (hero, FAQ, example prompts, platforms): https://cue.im
  • Phone numbers (Manus on X, 2026-09-28 17:43 UTC): https://x.com/ManusAI/status/2104628185501712789
  • "A Note to Our Users" (2026-08-11): https://manus.im/blog/a-note-to-our-users
  • Resumed independent operations (2026-09-01): https://manus.im/blog/manus-resumes-independent-operations
  • $100M ARR (2025-12-17): https://manus.im/blog/manus-100m-arr

Stripe

  • ✓ Link agent wallet FAQ (account created in the US or Canada only; biometric approval per purchase; one-time card): https://support.link.com/questions/link-agent-wallet-faq
  • Stripe customer story: https://stripe.com/customers/manus

Company

  • BBC (2026-04-27): https://www.bbc.co.uk/news/articles/cj0v0gr2yz7o
  • Computerworld (2026-08-14): https://www.computerworld.com/article/4209728/meta-gives-up-control-of-chinese-ai-startup-manus-after-eight-months.html
  • TechCrunch (2026-09-18): https://techcrunch.com/2026/09/18/manus-seeks-4b-valuation-in-new-500m-fundraise-as-it-resumes-independent-ops
  • The Decoder (2026-07-10), citing the FT: https://the-decoder.com/tencent-moves-to-buy-majority-stake-in-manus-after-beijing-forced-meta-to-unwind-its-2-billion-deal/

Launch press

  • InfoWorld (2026-09-29; Gartner and IDC quotes): https://www.infoworld.com/article/4228301/metas-ex-launches-agent-rival-to-metas-muse.html
  • The Decoder: https://the-decoder.com/manus-2-0-lets-users-edit-videos-host-multiplayer-games-and-run-agents-remotely-from-their-phone/
  • Thurrott: https://www.thurrott.com/a-i/342133/manus-2-0-arrives-with-a-new-architecture-and-impressive-new-capabilities
  • TestingCatalog: https://www.testingcatalog.com/manus-2-0-launches-with-studio-cloud-computer-and-cue/
  • Gigazine: https://gigazine.net/gsc_news/en/20260929-manus-2-0/
  • TokenPost: https://www.tokenpost.com/news/technology/24999
  • RuntimeWire: https://runtimewire.com/article/manus-cue-personal-agents-launch
  • CellCog (a competitor; launch timeline): https://cellcog.ai/blog/manus-2-0/

Security

  • Salt Labs, via OODA Loop (2026-09-25): https://oodaloop.com/briefs/cyber/prompt-injection-vulnerability-hits-4b-agentic-ai-platform-manus-ai/
  • Salt Labs, via Implicator (dateline 2026-09-28, 4:23 PM PST): https://www.implicator.ai/manus-cue-agents-phone-numbers-wallets/
  • CodeAnt AI (2026-09-19): https://codeant.ai/blogs/manus-prompt-injection-attack-ai-agent-security

Opinion (unverified; not relied on)

  • https://vocal.media/01/manus-just-handed-its-ai-agents-badges-its-own-founders-already-said-that-was-a-mistake

Lilith sources (approved revisions)

  • V1.3 PRD, Eve 8824516e-fae2-48e9-9e52-a1f938f5486a, revision 34: the comparison baseline, including §3 for Muse, dots, Instinct and Grok Bot, and every task named in sections 3, 4 and 7 except X10 and K11.
  • V1.3 PRD, the same document, revision 41 (approved; Eve allows one pin per document, so it is attached as an export with its content hash): the filed changes (status note).
  • V1.5 PRD, Eve b40979b3-f6f5-491e-a6f6-4564e7121268, revision 36 (Y06 code-as-shot, Z06, M4).
  • V1.4 PRD, Eve 549a0578-a216-4567-a298-1a58b780cce4, revision 20 (Z03).
  • The lilith-v1-3 board: X10, K11, the family counts and the dates the five new tasks were created (snapshot attached above).

The author's working notes are not attached; the captures above are the evidence for the claims they cover.