Agents (MCP)

Coding agents (Cursor, Claude Code, Claude Desktop, and other MCP clients) can read live project context and, on Team, write status back. They run as you, with your permissions. There is no bot user and no shared hive secret.

Hive (manager installs "Agents")
  └── You connect with your own PAT or OAuth
        └── Every tool call is your access

A member cannot use MCP until the hive has installed the plugin and they have connected under Account → Agents.

Plans

Capability	Plan
Read tasks, projects, knowledge	Studio+ (Ask tools also need Studio+)
Write status, comments, implement	Team
One-shot import from Linear / Jira / Asana / …	Every plan

Google Drive is org-wide OAuth. MCP is deliberately not. A shared “team MCP key” would bypass project membership and manager-only files for everyone at once.

1. Install (managers)

1. Confirm the hive is on Studio or higher (import-only tokens still work on every plan if Agent import is installed).
2. Open Organization settings → Plugins.
3. Install Agents (MCP). That only enables the feature. Nobody is connected yet.
4. Set the MCP write-mode policy (ceiling):
   • Confirm only (recommended) — members cannot turn on auto-apply.
   • Allow members to choose confirm or auto — Team only; each person still opts in under Account.

Then send people to Account → Agents. The Plugins card is install-only plus a link there.

2. Connect (every member)

Open Account → Agents from the avatar menu.

1. If you see an upgrade or “not installed” note, stop — a manager must finish step 1.
2. Create a personal access token. Name it after the client (Cursor, Claude Code, …) so Activity can attribute it.
3. Default scopes are read-only (tasks:read, projects:read, knowledge:read). On Team, tick Include write scopes if the agent should change tasks, comment, or file hive knowledge. Tick Include import:write only for a one-shot migration.
4. Copy the token now. It is shown once. Never commit it to git, a screenshot, or a shared chat.
5. Connect a client (below). Prefer OAuth when the client supports it — you will not paste a token.

Deactivating a person revokes all of their tokens. You can revoke one token any time; the agent stops until you mint another.

If a tool fails with a plan or permission error, that is your role and the hive’s platform plan — the agent cannot see more than you can in the UI.

Connect Cursor

Preferred: point Cursor at the MCP URL from Account → Agents without a PAT, then complete OAuth in the browser.

Fallback: copy the PAT snippet from the Cursor tab and paste it into Cursor’s MCP config. Keep X-InHive-Client: cursor so comments and activity show as Cursor, not a ghost edit.

The InHive Cursor pack (rules, skills, implement_from_brief) lives in the product repo under cursor-plugin/ until Marketplace publish lands. Account → Agents also shows the discovery URL (/.well-known/oauth-protected-resource on the same origin as /mcp).

Connect Claude Code

Copy the Claude Code snippet from Account → Agents into the project .mcp.json or your user Claude Code MCP settings:

• "type": "http"
• the hive’s MCP URL
• Bearer PAT (or OAuth if the client offers it)
• X-InHive-Client: claude

Connect Claude Desktop

No JSON. In Claude Desktop (or claude.ai): Settings → Connectors → Add custom connector. Paste the values from the Claude Desktop tab:

Field	What to paste
Name	InHive
URL	the remote MCP URL
OAuth Client ID	the value shown (advanced)
OAuth Client Secret	leave blank — InHive does not use one

Then Add, open the connector, and Connect. Sign in with your InHive email and password.

The URL must be a public https:// address. Claude connects from Anthropic’s cloud — localhost will not work from Claude Desktop.

Write mode

Account → Agents → MCP write mode:

• Confirm (default) — the client should ask before mutating tools.
• Auto — only on Team, only if the org policy allows it, and only if this token has write scopes.

Confirmation still happens in the client. InHive rate-limits, attributes the edit to you, and lets you revoke. This is a different gate from the in-app assistant, which uses approve/reject proposals.

Contract changes (acceptance criteria, the primary deliverable on a task, an accepted decision, filing hive knowledge) always become a proposal, even in auto. They show under From agents on the assistant and on Plan → Decisions until you or a manager approve them.

What agents are for

The wedge is client + brief + reviewed spec + project knowledge — not generic ticket sync. Point the agent at a project that already has material.

Empty knowledge makes the agent a substitute for memory. That is the failure mode. Add a brief or review a spec first. Hive procedures and information are readable with list_org_documents / get_org_document — search hits are short excerpts, not the full entry.

Copy prompt on a task ⋮ menu (Team, write access; enable it under Account → Profile) copies a local handoff the agent can pick up. Linking a pull request writes the branch / PR back onto the task. Set the project’s git remote in project settings so Cloud Agents can launch and so the agent can tell when it has the wrong client repo open.

Start a project with get_project_charter — purpose, dates, who at the client can accept, goals, included and excluded deliverables, settled constraints, and the loudest open unknowns. Fetch get_project_rules for the accepted Decision Log and reviewed specs as a rules file (rules_version tells you when to refresh). get_context_package and get_handoff put that charter first. Pass compact: true when you only need id+title refs; fetch bodies with search_knowledge.

After create_task, update_task, or link_pr, InHive may flag a contradiction against an accepted decision. The write still succeeds. The Decision Log shows a conflict badge; a manager or the decision owner can dismiss it. Changing a task title or description in a way that looks like a scope change can be held as a proposal even in auto mode — acceptance-criteria edits already were.

When an agent marks a task done, that completion is Declared until a human who did not do the work verifies it. The agent sees verification on the task and a note on the status reply. Agents cannot verify. See Tasks. get_task, list_my_work, and list_project_tasks also include blocked_by and ready so a blocked task is visible without opening the board. list_project_tasks includes each task’s label (name and color) when one is set. get_task_history is the same History of changes as the task drawer — who moved it to QA and when — not just the current status. Comment text is a separate call, get_task_discussion: the newest comments, and the distilled brief if someone has already saved one. That call does not distill the task for you. A stale brief means the description or a comment changed after it was saved.

MCP calls count against hive AI usage and a per-person rate limit. A looping agent will burn spend.

Project context (tasks, briefs, reviewed specs, knowledge) may leave InHive and enter the agent’s context. Use MCP only on client work where that is allowed.

Import from another tool

Agent-driven import (Linear, Jira, Asana, …) is a one-shot session: push items, preview, commit. Available on every plan. It is not a live sync. An import-only token cannot read the rest of the hive.

Monday has its own first-party wizard — prefer Plugins for Monday boards.

Related

AI settings · Knowledge · Plugins · Tasks