How do I let my own Claude Code session investigate crash loops with a project key?

Tares lets your own Claude Code session investigate crash loops through a project key that reads one project and records findings only there. You create the key under Settings, Keys, copy the secret once, and connect with claude mcp add and an Authorization Bearer header. Claude Code then reads the project’s sources and posts its conclusion next to what the Tares agents already found.

How it works

A project key ties one MCP connection to one Tares project. The session you start with it is an outside agent: it is not run by Tares, but it works on the same data.

• The key is scoped to the project. Any call outside that project answers 403.
• The project is implied, so the tools need no project argument.
• The key’s name is stamped on every finding it records, so you can tell your session from the Tares agents.
• Claude Code reaches Tares through the MCP endpoint that tares mcp serves at http://127.0.0.1:8788/mcp.
• The key is checked only when the instance runs with auth, either tares up --auth or TARES_AUTH_TOKEN set. On an open instance there are no keys, so you skip the key and name the project in your prompts.

For a restart problem, the session uses the project tools in order. It loads a matching skill with list_skills and get_skill. It calls stats to see which values moved and read for the affected entity’s timeline across the project’s sources. It checks list_findings and project_timeline for earlier conclusions. When it has an answer it calls record_finding(entity, finding, verdict). The entity is what the finding is about, the finding is markdown, and the verdict is one lowercase word such as rca.

Set it up

1. Make sure the project already has the sources that show the failing service, and that the instance runs with auth and the MCP endpoint (tares mcp).
2. Open the project, go to Settings → Keys and choose New key.
3. Name the key after the agent that will hold it, for example claude-code. Copy the secret straight away, because it is shown once.
4. Connect Claude Code with the key:

claude mcp add --transport http tares http://localhost:8788/mcp \
  --header "Authorization: Bearer <project key>"

5. If you run the MCP server over stdio instead, start tares-mcp with TARES_AUTH_TOKEN=<project key> in its environment.
6. Check the link from a Claude Code session:

Use tares: which project can you work in?

It calls list_projects and names the single project your key belongs to. 7. Point the session at the restarting service:

Use tares: api-server keeps restarting. What changed in the last 30 minutes? Check the project’s skills first, and record what you conclude.

Swap api-server for your own entity. The outside agent guide covers every step in full, and the AI SRE demo gives you a project with sources and a trigger to practice on.

What you get

• Your session investigates the crash loop with the project’s own sources and skills, not a blank context.
• It builds on earlier Tares agent findings instead of repeating them.
• Its conclusion appears as a card under What the agents found on the project’s Overview, under the key’s name.
• In Advanced setup’s Activity the finding is a thread of its own, marked as recorded from outside Tares.
• The finding sits on the entity’s timeline in Explore, so the next Tares agent that reads that service starts from your conclusion.
• Cutting access is one action: revoke the key under Settings → Keys.

When something else fits

• If you want the agent to run without you asking, use join_project(url). Every trigger firing in the project is sent to that URL, and the outside agent shows under Agents marked external.
• For Codex, Cursor or your own agent, the steps are the same and only the connect command differs. The outside agent guide points to the full list of clients.
• If you only want the Tares agents to investigate, you do not need a key.

Related