The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Owl24 MCP listing page.
MCP server for owl24 — hand a production error to the coding agent you already use (Claude Code, Cursor, or any MCP-speaking client). It claims the item, reads the trace and stack trace, and writes the fix — instead of you reading a stack trace for a file you've never opened.
This doesn't replace the REST-plus-AGENTS.md workflow — it's a thinner way to reach the exact same endpoints. If you already have an AGENTS.md-driven CI loop working, you don't need this. This is for working with your agent interactively, from inside your editor.
| Tool | Does |
|---|---|
list_errors | Deduplicated error groups — occurrence counts, first/last seen, queue state, whether an AI report already exists |
queue_error | Adds an error group to the agent-access queue (free, never calls AI) |
list_queue | Lists queue items — defaults to open/unclaimed |
claim_error | Claims a queue item so nothing else works it at the same time — this is what's billed, once (see below) |
get_trace | The actual evidence — spans, logs, stack trace for one trace |
get_rca_report | Checks for an existing cached AI root-cause report (read-only) |
resolve_error | Marks a queue item resolved, with a note naming the PR and fingerprint — free, doesn't trigger any charge |
count_open_items | Cheap poll target — just the open-item count, no full list fetch |
release_error | Gives up on a claimed item without resolving it — doesn't add a new charge (a claim's charge, if any, already happened) |
extend_claim | Pushes a claim's expiry out if you're still actively working it |
What it deliberately can't do: run a new AI Root-Cause Analysis. That's a separate, paid, human action from the owl24 dashboard — this server never spends AI-RCA credit on its own initiative. get_rca_report only reads a report that already exists; if one doesn't, investigate from get_trace's data instead, the same way AGENTS.md already tells an agent to.
Billing: listing and queuing are always free. A small charge (current rate on Pricing) applies once, the moment claim_error succeeds on an item that's never been claimed before — not when it's later resolved, and with no monthly cap. If you release_error instead of fixing it, or the claim simply expires, that charge isn't refunded, but re-claiming the same item later is never billed again. resolve_error itself never adds a charge — use it once you actually have a fix, and use release_error (also free to call) when you don't, so the record stays accurate for whoever claims it next.
We don't build, run, or host your agent. This server runs locally, on your machine, using your own project API key — same trust model as the REST workflow it wraps.
You don't need to install this yourself — most MCP clients run it via npx on demand.
Add to your MCP settings (Cursor Settings → MCP):
Any other MCP client: the same command/args/env shape works — npx -y owl24-mcp over stdio, with OWL24_API_KEY in the environment.
Get your API key from a project's page on the owl24 dashboard. Agent Access needs to be turned on for that project first, via your account's own API — every tool here 403s until it is.
| Variable | Required | Default |
|---|---|---|
OWL24_API_KEY | Yes | — |
OWL24_API_BASE | No | https://api.owl24.dev — override for self-hosted/local testing |
Once connected, in Claude Code or Cursor:
What's breaking in checkout-service?
The model calls list_errors({ serviceName: "checkout-service" }), reads the results, and can go straight into get_trace for the top offender, write the fix itself, and — once you've opened the PR — resolve_error with a note. A human always reviews the PR; nothing here merges anything.
MIT