Asana AI助手集成

MCP Server for Asana

Give your AI assistant full access to Asana. Create tasks, manage projects, analyze delivery flow, estimate work from history — all through natural language.

86 tools | 35 slash commands | Autocomplete | Browsable resources | Full read + write coverage

What can you do with this?

Just talk to your AI assistant naturally:

"How many unfinished tasks are in Sprint 30?"

"Create a task for Sarah to review the API docs, due Friday, in the Backend project"

"Why is this task late?" → full lifecycle: where it stalled, who touched it, every reschedule

"What did Alice work on in the last 3 months?"

"How much should we commit to next sprint?" → measured from real velocity, not guesswork

"Where does work get stuck in this project?" → cycle time, bottlenecks, throughput

"Turn these meeting notes into tasks"

Run /asana-help to see every command grouped by what you're trying to do.

Highlights

Quick Start

1. Get your Asana token from Asana Developer Console

2. Add the server to your MCP client:

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "asana": {
      "command": "npx",
      "args": ["-y", "@blzvi/asana-mcp-server"],
      "env": {
        "ASANA_ACCESS_TOKEN": "your-asana-access-token"
      }
    }
  }
}

Claude Code

claude mcp add asana -e ASANA_ACCESS_TOKEN=<TOKEN> -- npx -y @blzvi/asana-mcp-server

Any MCP-compatible client

This server works with any client that supports the Model Context Protocol:

ASANA_ACCESS_TOKEN=your-token npx -y @blzvi/asana-mcp-server

Install from source (alternative)

git clone https://github.com/BLZvi/asana-mcp-server.git
cd asana-mcp-server
npm install

Then use node build/index.js instead of npx -y @blzvi/asana-mcp-server in the examples above.

86 Tools Across 20 Categories

Full CRUD coverage of the Asana API — your AI can read and write.

35 Slash Commands

Pre-built workflows that combine API calls with AI reasoning. The server pre-fetches the relevant Asana data server-side, then hands the model one fully-hydrated message — so most commands need zero follow-up tool calls.

Not sure which to use? Run /asana-help. ✎ = writes to Asana (always previews first).

Task-level

Project-level

Personal productivity

Cross-project

Planning

Code ↔ Asana

These need your client to have git/filesystem tools (Claude Code, Cursor, etc). They degrade gracefully when it doesn't — asking you to paste the command output instead.

Discovery

Browsable Resources

Expose Asana data as MCP resources that clients can browse directly:

Configuration

Bad or missing config never throws — every value degrades to its default.

Set ASANA_TIMEZONE. The default of UTC means a user in UTC-8 at 5pm gets tomorrow's date, and tasks get flagged overdue a day early.

Safety Modes

Read-only (ASANA_READ_ONLY_MODE=true) — hides all 43 write tools and 7 write commands. For safe exploration, shared environments, and audit/reporting use.

Dry run (ASANA_DRY_RUN=true) — write tools stay visible so the model can plan a complete sequence, but each returns "would have created task X" instead of calling Asana. Use this to rehearse a bulk operation before committing to it.

Tool annotations — every tool reports readOnlyHint, destructiveHint, and idempotentHint, so clients can auto-approve safe reads while prompting for confirmation on the 13 genuinely destructive operations.

How the analytics stay honest

Asana's task search endpoint returns at most 100 results and supports no pagination. A naive "what did Alice do in 6 months" query returns an arbitrary 100 tasks and reports it as complete.

This server solves that with adaptive time-window bisection: the range is split into buckets, and any bucket that saturates at 100 is recursively halved until it fits or hits a floor. Results are deduplicated by GID.

Critically, every command using this reports its own coverage:

Based on 340 tasks across 7 windows (28 requests). Complete.

or

Based on 512 tasks across 12 windows (40 requests). INCOMPLETE — 2 windows hit the 100-result cap; counts are a lower bound.

The same principle applies throughout: flow-report states its sample size versus project size, contributions ships methodology caveats with the data, and sprint-capacity suppresses point velocity entirely when fewer than 50% of tasks have points set — because low-coverage point velocity looks authoritative while measuring an arbitrary subset.

Roadmap

What's coming next:

• Elicitation — Structured confirmation dialogs for bulk writes (currently prose-based preview + confirm)
• Structured output — outputSchema on read tools so clients render tables instead of JSON
• Remove dependencies/dependents — Complete dependency management (currently add-only)
• Story CRUD — Edit and delete comments (currently read + create)
• Duplicate project — Create projects from templates
• Test coverage — Unit and integration tests with vitest

Full Tool Reference

Security

This server holds an Asana personal access token with your full permissions. It writes nothing to disk, talks only to app.asana.com, and never logs credentials.

That last point takes active effort: the Asana SDK attaches the Authorization: Bearer <token> header to every error it throws, so a naive console.error(err) would leak your token into terminal scrollback and client logs. All error output is routed through a redacting summariser, and CI fails if any code path passes a raw error to console.*.

See SECURITY.md for the full threat model and reporting process.

Requirements

• Node.js 22 or newer. CI tests against 22 (Maintenance LTS), 24 (Active LTS), and 26 (Current). Node 24 is recommended — 22 leaves maintenance in April 2027.
• An Asana personal access token

Contributing

git clone https://github.com/BLZvi/asana-mcp-server.git
cd asana-mcp-server
npm install
npm run verify   # lint, typecheck, tests, logging guard, build, smoke test

Then npm run dev to run from source, or npm run inspector for interactive testing.

See CONTRIBUTING.md — it documents four rules that are enforced in CI (never log raw errors, never hide incomplete coverage, never parse story text, always preview before writing).

License

MIT