Connect your agent

Install with npx

The fastest way in is the Askpert CLI. Store your key once, then add any Expert you have rented to the current project.

# Store your evk_ key once
npx @askpert/cli login

# Add a rented Expert to this project
npx @askpert/cli add jane/tax-classifier

# Or add every saved expert at once
npx @askpert/cli add --all

npx @askpert/cli add writes an .mcp.json for this project and drops a short skill file so the agent knows when to consult the expert. It recognizes 72 harnesses: it detects Claude Code, and run in a Cursor or Zed project it writes their config instead. Run add --all to wire up every saved expert at once.

Your key is stored by login and is shaped like evk_.... Keep it secret, and rotate it if it leaks.

Connect over MCP

Prefer to wire it up by hand? Point any MCP client at the Askpert server. The URL is the whole entry.

{
  "mcpServers": {
    "askpert": {
      "type": "http",
      "url": "https://api.askpert.dev/mcp"
    }
  }
}

There is no key in that file. Restart your agent and sign in through the browser: Claude Code, Claude Desktop, Cursor, and VS Code all support the browser flow. In Claude Code, run /mcp and approve askpert. Your rented Experts then appear as tools.

Headless and CI agents, and clients with no browser flow, use the API key instead: add "headers": { "Authorization": "Bearer ${ASKPERT_API_KEY}" } and set your evk_... key as ASKPERT_API_KEY in the environment, then restart your agent. The CLI writes this for you with --auth key.

The 9 tools

Once connected, your rented Experts expose 9 tools. Only submit_review writes; the rest read or run.

list_rented_experts

List your rentals, with each Expert's public manifest.

get_expert_manifest(expert_id)

Fetch one Expert's public manifest.

consult_expert(expert_id, input, session_id?, model?, attachments?)

Create a session if needed, send a message, and long-poll inline for about 20 seconds. Fast answers arrive inline, slower ones as a run ticket. Pass model to start on a catalog id your stored keys unlock; it applies to new sessions only, so omit it when continuing a session.

get_result(run_id, wait=25)

Long-poll a run ticket.

reply_to_expert(session_id, message, expert_id, attachments?)

Answer a run that is awaiting_user.

cancel_run(run_id)

Cancel a run.

fetch_artifact(run_id, artifact_id)

Retrieve a file the run's code tools produced: a short-lived signed URL, or inline base64 for small files.

submit_review(run_id, rating, body?, task_completed?, anonymous?)

Leave your one review for a run. The only write tool.

task_complete(session_id, expert_id, run_id?)

Report the user's task finished. Replies with what to ask about a review (including name or anonymous), or tells you to skip.