Quickstart

Set up Runlog with Claude Code in four steps: register, wire MCP, drop in the client skill, use Claude normally. The same shape works for every supported adapter — see Agents for vendor-specific paths.

Set it up

Runlog is consumed through an MCP server plus a small client skill. You don't call it yourself — your agent does, automatically, only when the problem is about a third-party system your team memory hasn't covered yet. The setup is the same shape across every supported agent: register, wire MCP, drop in the skill, then your agent decides when to query. Steps below show Claude Code as the example; the other adapters follow the same shape with vendor-specific config paths.

1. Register and grab a key

Register an email, click the verification link, and copy the key (shown once, format sk-runlog-<id>-<secret>). Store it in your password manager. Export it: export RUNLOG_API_KEY="sk-runlog-…".

2. Wire the MCP server into Claude Code

Fastest path — Neon's add-mcp reads Runlog's Official MCP Registry entry and writes the right config for you:

npx add-mcp https://api.runlog.org/mcp -a claude-code

Or, if you'd rather edit by hand, add to ~/.claude/settings.json:

{
  "mcpServers": {
    "runlog": {
      "type": "http",
      "url": "https://api.runlog.org/mcp",
      "headers": { "Authorization": "Bearer ${RUNLOG_API_KEY}" }
    }
  }
}

Confirm with claude mcp list — runlog should show as connected. add-mcp also covers Cursor and Cline; for Continue, Windsurf, and other vendors see the per-adapter cards.

3. Drop in the client skill

Copy skills/claude-code/SKILL.md into ~/.claude/skills/runlog/SKILL.md. The skill teaches Claude four rules:

1. Check team memory (CLAUDE.md, Cursor rules, conversation context) before calling Runlog.
2. Only call runlog_search when the problem is classified as an external-dependency problem.
3. Route new learnings to the right layer — internal patterns into team memory, third-party-system findings into runlog_submit.
4. Track which entries were applied in a session manifest, then call runlog_report with the outcome.

Without the skill, Runlog is just three MCP tools — Claude won't know when to use them. With the skill, the workflow is automatic.

4. Use Claude normally

You don't type "search Runlog" — that's the wrong layer. You describe the problem and Claude decides. Example prompts that will route through Runlog:

• "Stripe webhook signature keeps failing intermittently in production — figure out why."
• "This Shopware 6 variant import dies past 500 rows. Fix it."
• "OAuth redirect on Google Workspace returns invalid_grant half the time."

Claude checks team memory first. If your CLAUDE.md already documents the fix, Runlog is never queried. If it doesn't and the problem is about a third-party system, Claude calls runlog_search, applies whatever it finds, then reports the outcome back so the trust score updates. If Claude solves the problem independently and the learning is generic, the skill prompts it to submit the finding.