Owl24 MCP

What owl24-mcp does

The owl24-mcp MCP server gives an MCP-compatible coding agent access to production errors tracked by owl24. An agent can find deduplicated error groups, inspect their queue state, and select an item to work on. After claiming an item, it can retrieve the underlying trace data, including spans, logs, and stack traces, then make code changes in the connected development workflow.

The server also supports queue housekeeping. Agents can add an error to the access queue, list open or claimed work, check the open-item count, extend an active claim, release work that will not be resolved, or mark an item resolved with a note. Resolution notes can identify the relevant pull request and error fingerprint.

How it works

The process is designed for interactive agent use rather than autonomous remediation. A typical request causes the agent to call list_errors, choose an error group, inspect it with get_trace, and work on the fix. Once a pull request has been opened, the agent can call resolve_error with a note. Nothing in the server merges code, and a human reviews the pull request.

get_rca_report only reads an existing cached owl24 AI root-cause report. The owl24-mcp MCP server does not start a new RCA job; a new analysis is a separate paid action performed by a person from the owl24 dashboard. If no report exists, the trace returned by get_trace is the evidence available to the agent.

Claiming work has a billing rule that differs from the other queue actions. A successful first claim for an item incurs the applicable charge once. Releasing the item or allowing the claim to expire does not refund that charge, but claiming the same item again is not billed a second time. Listing, queuing, resolving, releasing, and extending claims do not add a charge.

Setup and configuration

Run the package locally through an MCP client's stdio configuration. The documented command is npx -y owl24-mcp; most clients can start it on demand without a separate installation. Set OWL24_API_KEY to a project API key from the owl24 dashboard. Agent Access must be enabled for that project, otherwise the tools return HTTP 403 responses.

OWL24_API_BASE is optional and defaults to https://api.owl24.dev. Set it when using a self-hosted or local testing API. Claude Code can register the server with claude mcp add, while Cursor accepts the same command, arguments, and environment values in its MCP settings. Other stdio-capable MCP clients use the same configuration shape.

Tools and capabilities

The available tools are:

• list_errors: Returns deduplicated error groups with occurrence counts, timestamps, queue state, and report availability.
• queue_error: Adds an error group to the agent-access queue without calling AI.
• list_queue: Lists queue items, defaulting to open and unclaimed entries.
• claim_error: Claims an item so concurrent agents do not work on it simultaneously.
• get_trace: Retrieves spans, logs, and a stack trace for a trace.
• get_rca_report: Reads a cached RCA report when one already exists.
• resolve_error: Records resolution and a note without triggering another charge.
• count_open_items: Returns the number of open items for inexpensive polling.
• release_error: Ends a claim without resolving the item.
• extend_claim: Extends the expiry of an active claim.

Limitations and notes

The owl24-mcp MCP server does not build, run, or host the coding agent. It also does not initiate new AI RCA work, merge pull requests, or remove the need for human review. It is a thinner interactive route to owl24's existing endpoints, not a replacement for an AGENTS.md-driven CI workflow. Projects without Agent Access enabled cannot use the tools successfully.