Connect an AI assistant

What it lets you do

Once connected, your assistant is talking to your actual workspace. So you can ask it things only your board knows the answer to:

• —“What’s on my plate this week, and what’s overdue?”
• —“Which cards on the Roadmap board are blocked, and what’s blocking them?”
• —“Add a card to Roadmap for the billing migration, high priority, due Friday.”
• —“Summarise the comments on TAR-142 and tell me what was decided.”
• —“I was off for a week. What actually changed?”

It reads and writes the same board you use. A card it creates is a real card, and a comment it posts is attributed to you and notifies your team.

Set it up

1. 01

   Create an API key

   In Tarsen, go to Workspace → Integrations → API keys. Give it a name (“Claude Desktop”, say) and press Create key. Copy it straight away, because for security we only store a scrambled version, so it can never be shown again.

2. 02

   Find your workspace address

   It is your Convex site URL with /mcp on the end. If you are self-hosting or unsure, it is the NEXT_PUBLIC_CONVEX_SITE_URL value from your deployment, and it ends in .convex.site.

   https://your-deployment.convex.site/mcp

3. 03

   Add it to Claude Code

   claude mcp add --transport http tarsen \
     https://your-deployment.convex.site/mcp \
     --header "Authorization: Bearer tsn_your_key_here"

4. 04

   Or add it to Claude Desktop

   Settings → Developer → Edit Config, then paste this in and restart the app.

   {
     "mcpServers": {
       "tarsen": {
         "type": "http",
         "url": "https://your-deployment.convex.site/mcp",
         "headers": {
           "Authorization": "Bearer tsn_your_key_here"
         }
       }
     }
   }

Any other client works too, as long as it supports HTTP with custom headers. The server speaks standard JSON-RPC and the key goes in an Authorization header.

Shortcut: right after you create a key, press “Copy agent prompt” on that same screen. It copies a message with your real server URL and key already filled in — paste it straight into your agent's chat and it can set the connection up itself.

What your assistant can reach

14 tools, covering projects, boards, cards and comments. 5 only read; 9 can change something, and 6 of those are held to a scoped key's own work (see “Is this safe?” below).

Is this safe?

A key acts as you, in one workspace. That is the whole rule, and everything else follows from it:

• —It reaches exactly the projects you can reach. A private project you are not a member of is invisible to your assistant, just as it is to you.
• —It cannot do anything you cannot. If you have view-only access to a project, so does your key.
• —It is locked to the one workspace it was made in. There is no way to point it at another.
• —If you leave the workspace, the key stops working immediately.
• —Keys are personal. Even a workspace admin cannot see or use yours.

There is a second, optional layer for agent keys: when you create a key, you can check “Only let this agent change work assigned to it”. A scoped key can still read the whole board and claim any unassigned card, but it cannot move, relabel, comment on or reassign a card held by someone else — including another agent. This is what stops one runaway agent from touching a teammate's work while it is left unattended.

You can revoke a key at any time from the same screen you created it on. It stops working on the next request.

Because a key can create cards and post comments, treat it like a password. If you paste one somewhere public, revoke it and make a new one. It takes about ten seconds.

If something goes wrong