Site
- URL: https://manus.im (entered via
/app)
- Auth: standard cookie-mode login
- Stack: Next.js (chunks served from
files.manuscdn.com/webapp/_next/...)
- API:
/api/* (disallowed in robots.txt — needs logged-in session)
- Public surfaces:
/share/<id>, /report/<id>, /invitation/<id>
Why it fits OpenCLI
Manus is positioned as an "action engine" — users submit task prompts, the agent runs autonomously for minutes-to-hours, and produces step traces + deliverable files (code/docs/datasets). It maps cleanly onto OpenCLI's existing AI-surface patterns:
- Read-side commands look like
kimi / qoder adapters (status / list / read / history)
- Task submission is similar to
codex new-task / antigravity patterns (--yes gated)
- Long-running agent execution suggests
manus status <id> polling and manus read <id> --include-steps for deliverable retrieval
Proposed command surface (subject to DOM verification)
Read-only / idempotent — should be safe under the existing decision tree:
manus status — current account + active sessionmanus list — task / conversation listmanus read <id> — task detail + step trace + deliverables URLsmanus history — recent N tasksmanus open <id> — navigate to task page (idempotent)manus storage-keys / storage-get — localStorage introspection
Destructive / state-changing (only if a confirmation dialog or verifiable post-state exists):
manus new-task <prompt> --yes — submit a new taskmanus chat <id> <message> — append message to running taskmanus pause / resume / cancel <id>
Plan
If this is welcome, I'll do the same shape as the recent Kimi (#1804) and Qoder (#1803) PRs:
- vitest helper tests
docs/adapters/<bucket>/manus.mdcli-manifest.json rebuildnpm test green locally
Let me know if you'd rather see a smaller subset first (e.g. read-only only) before submission picks up speed.
Site
/app)files.manuscdn.com/webapp/_next/...)/api/*(disallowed in robots.txt — needs logged-in session)/share/<id>,/report/<id>,/invitation/<id>Why it fits OpenCLI
Manus is positioned as an "action engine" — users submit task prompts, the agent runs autonomously for minutes-to-hours, and produces step traces + deliverable files (code/docs/datasets). It maps cleanly onto OpenCLI's existing AI-surface patterns:
kimi/qoderadapters (status / list / read / history)codex new-task/antigravitypatterns (--yesgated)manus status <id>polling andmanus read <id> --include-stepsfor deliverable retrievalProposed command surface (subject to DOM verification)
Read-only / idempotent — should be safe under the existing decision tree:
manus status— current account + active sessionmanus list— task / conversation listmanus read <id>— task detail + step trace + deliverables URLsmanus history— recent N tasksmanus open <id>— navigate to task page (idempotent)manus storage-keys/storage-get— localStorage introspectionDestructive / state-changing (only if a confirmation dialog or verifiable post-state exists):
manus new-task <prompt> --yes— submit a new taskmanus chat <id> <message>— append message to running taskmanus pause / resume / cancel <id>Plan
If this is welcome, I'll do the same shape as the recent Kimi (#1804) and Qoder (#1803) PRs:
docs/adapters/<bucket>/manus.mdcli-manifest.jsonrebuildnpm testgreen locallyLet me know if you'd rather see a smaller subset first (e.g. read-only only) before submission picks up speed.