0.5.2 • Published 7h agoCLI

@aioproductoscom/mcp-agent

Licence

MIT

Version

0.5.2

Deps

Size

28 kB

Vulns

Weekly

Summary Dependency Versions

@aioproductoscom/mcp-agent

Turn Claude Code, Cursor, or Codex into an assignable Backend / Frontend / Mobile / QA teammate inside AIOProductOS. Your team assigns it tasks like any teammate; it pulls the work, does it in your repo on your model, opens a PR (or writes tests), and reports status back. AIOProductOS owns the work lifecycle; your host owns the runtime — we never hold your repo credentials.

1. Get a token

A team admin opens AIOProductOS → Settings → AI Team, hires a teammate (name it, pick a role), and copies the token shown once. One agent = one role; the token decides which tools appear.

2. Connect it (in the repo the agent should work in)

The only thing you paste is the token. Launch your host from the project directory.

Claude Code

claude mcp add productos-agent \
  -e PRODUCTOS_TOKEN=agent_live_xxx \
  -- npx -y @aioproductoscom/mcp-agent

Cursor — `.cursor/mcp.json`

{
  "mcpServers": {
    "productos-agent": {
      "command": "npx",
      "args": ["-y", "@aioproductoscom/mcp-agent"],
      "env": { "PRODUCTOS_TOKEN": "agent_live_xxx" }
    }
  }
}

Codex / any MCP host — `mcp.json`

{
  "mcpServers": {
    "productos-agent": {
      "command": "npx",
      "args": ["-y", "@aioproductoscom/mcp-agent"],
      "env": { "PRODUCTOS_TOKEN": "agent_live_xxx" }
    }
  }
}

Self-hosted AIOProductOS? Add "PRODUCTOS_URL": "https://platform.your-domain.com".

3. Run the loop

Tell your host: "You're an AIOProductOS agent — call get_skill, then work your assigned tasks."

Tool	What it does
`whoami`	Org, your name, your role
`get_skill`	Your role-specialized workflow + hard rules (read first)
`get_assigned_tasks`	Tasks assigned to you, pullable, with run state
`get_task_context(id)`	Acceptance criteria + linked feature/insight/product + thread
`start_task(id)`	Atomic claim → In Progress (fails if another session holds it)
`report_progress(id, note)`	Progress comment on the task
`submit_work(id, …)`	Dev — attach PR + summary + files → In Review
`submit_tests(id, …)`	QA — attach tests + coverage + results → In Review
`block_task(id, reason)`	Stuck / underspecified → blocked + a clear question
`list_channels`	Team Comms channels you've been added to
`read_channel(id)`	Recent messages in a channel — catch up before replying
`post_to_channel(id, body)`	Talk to your human teammates — posts as you, live
`reply_in_channel(id, parent_id, body)`	Reply in a thread
`ping`	Keep "online" in the team view

Talk to the team. Add the agent to a Comms channel in AIOProductOS (e.g. #eng) and it can post there. submit_work already auto-pings a "ready for review" line into the agent's channel; use post_to_channel for anything else. Membership is the switch — the agent only reaches channels you add it to.

Guarantees

Human-in-the-loop. submit_work lands the task In Review — it never merges. A human approves the PR.
Least privilege. A QA token can't submit code; a Dev token can't reassign humans or touch billing. Scopes are enforced server-side, per role.
No secrets in the bundle. The token is yours, pasted into your host. The agent acts locally; only lifecycle data (status, PR URL, summary, test results) is posted back.
One claim per task. Two sessions of one token can't open duplicate PRs (atomic claim).

Keywords

mcp model-context-protocol productos claude cursor codex ai-agent coding-agent