MCP setup

Use Done in Claude Code

Add Done to Claude Code with one claude mcp add command, sign in from /mcp, and let Claude Code pick up, plan and report back on your Done tasks.

Who this is for

For developers using Claude Code in a terminal or IDE. It fits when the work on your list is code: Claude Code can claim a task, propose a plan, do the work in your repository and send it back to you for approval.

Set it up

Done's MCP server URL

https://trydone.app/mcp

Add the server and sign in Sign in with OAuth

Run this once in a terminal. --scope user makes Done available in every project; leave it off to add Done to the current project only.
```
claude mcp add --transport http --scope user done https://trydone.app/mcp
```
Start Claude Code, run /mcp, choose done, and follow the sign-in in your browser. From a shell, claude mcp login done does the same.
Approve access in Done. Claude Code stores the sign-in and refreshes it on its own.

Or use a token MCP token

In Done, open Settings → Agent access and generate an MCP token. It's shown once, so copy it straight away.

Add the server with your token as a header:

claude mcp add --transport http --scope user done https://trydone.app/mcp \
  --header "Authorization: Bearer <your-mcp-token>"

Don't add a server that carries a token with --scope project: that writes it into .mcp.json, which usually gets committed.

Need a narrower key? Under Settings → Agent access → Extra MCP tokens you can create one that's read-only, or limited to one project or one workspace — the agent then only ever sees that much.

What to say to it

Once Done is connected, talk to the assistant normally. Each of these maps to a real Done tool:

“What's on my plate today?”

get_todays_plan

“Add a task: send the invoice to Sam friday p1”

quick_capture

“Pick up your next Done task and propose a plan before you start.”

next_task → propose_plan

“If you hit a decision you can't make, ask me and keep going with the rest.”

ask_decision

“When it's finished, send it to me for approval.”

propose_completion

Every tool, prompt and resource Done exposes is listed in the MCP reference.

The approval gate

By default an assistant can't close your tasks. When it thinks a task is finished it calls propose_completion, which parks the task as Pending approval until you approve it or send it back from Needs you; substantial work starts with propose_plan, which waits for your go-ahead. A question goes through ask_decision: that task waits for your answer while the agent moves on to the next one. Done enforces this on its own server, so it holds whatever the assistant tries — an agent can only complete a task directly when you've set that task's Autonomy to let it.

Troubleshooting

done shows as needing authentication: Run /mcp, select done and sign in. To start over, choose Clear authentication there and sign in again.
The browser didn't open: Copy the sign-in URL Claude Code prints and open it yourself.
A token that used to work now gets 401: Generating a new primary token revokes the old one. Create a new token, then claude mcp remove done and add the server again.

FAQ

Can several Claude Code sessions share my queue?

Yes. next_task claims a task atomically, so two sessions never pick up the same one. Give each its own named token to see their work separately on your roster.

Is this the same as Claude on the web?

No — that's a custom connector, set up in Claude's settings. See Use Done in Claude.

Is there a shared API key for Done?

No. Every connection belongs to one Done account: an OAuth sign-in, or an MCP token you generate yourself. There is no shared or global key. Each connected assistant is listed under Settings → Agent access, where you can revoke it in a click.

What can the assistant see?

Only what your account can see: your workspaces and anything shared with you. A read-only, one-project or one-workspace token narrows that further.

Does it cost anything?

The Free plan includes 100 agent (MCP) calls a month; Pro and Teams are unlimited. The assistant's own plan is separate.

Steps checked against the official documentation on October 23, 2026:

code.claude.com/docs/en/mcp