@error-bar/mcp

@error-bar/mcp

The errorbar platform as MCP tools — every public API operation (evals, judges/criteria, gates, logs, datasets, aliases, label sets, audit/proving, dedicated endpoints, training) callable from Claude Code, Claude Desktop, Cursor, or any MCP client.

One tool per API operation, 76 tools, generated from the platform's route contracts. Inputs are validated with the same field names and casing the REST API accepts; responses come back exactly as the API returned them (JSON pretty-printed, exports as text), including the API's own error status and message.

Renamed from @omnia-voice/mcp (2026-08-30). The old name is published as a shim that forwards here; update your config when convenient.

Install

npx -y @error-bar/mcp        # or: npm i -g @error-bar/mcp && errorbar-mcp

Needs ERRORBAR_API_KEY (a workspace API key, sk_…, from Settings → API keys). Give the key only the scopes the agent should have — read for a monitoring agent, read + evals:write for one that runs evals — the platform enforces them and each tool's description says which scope it needs.

Claude Code

claude mcp add errorbar -e ERRORBAR_API_KEY=sk_… -- npx -y @error-bar/mcp

Claude Desktop / Cursor / any mcpServers config

{
  "mcpServers": {
    "errorbar": {
      "command": "npx",
      "args": ["-y", "@error-bar/mcp"],
      "env": { "ERRORBAR_API_KEY": "sk_…" }
    }
  }
}

Profiles — hand the agent the tools for one job

Seventy-six tools is the complete contract and the wrong thing to give a model with one question: every description is read before the first call, and a long list makes the pick worse. A profile is the handful a job needs, plus the task-shaped tools that answer the job's question directly.

npx -y @error-bar/mcp --profile monitor      # or ERRORBAR_MCP_PROFILE=monitor

--read-only, --no-spend and --only apply on top of a profile.

Task-shaped tools

Three tools answer the questions people actually ask, with the doctrine applied (verdicts from the interval, never the point estimate) and a pointer to the one-per-endpoint tool for the detail:

Resources and prompts

Resources: errorbar://agent-setup.md and errorbar://llms.txt (fetched live from www.errorbar.ai, so an installed server never rots) and errorbar://after-traffic-flows (the first five steps, each naming the tool that performs it). Prompts: set-up-errorbar, after-traffic-flows, release-check (eval_id optional).

Flags

The OMNIA_* spellings of every environment variable are still accepted.

Tools that spend wallet credit say SPENDS MONEY in their description and are not readOnlyHint; MCP clients that confirm before non-read-only calls will ask for those.

Tools

Each tool's description carries the full contract: every query and body field, what the response contains, and the gotchas (feature flags, 400 conditions, cursor semantics).

Programmatic use

import { createServer, ErrorbarClient } from "@error-bar/mcp";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";

const server = createServer({ client: new ErrorbarClient({ apiKey: process.env.ERRORBAR_API_KEY! }), readOnly: true });
await server.connect(new StdioServerTransport());

Development

npm ci && npm run typecheck && npm test && npm run build

src/manifest.ts is generated from the platform's app/api/v1 routes; test/manifest.test.ts pins the exact route inventory it must cover, so a new platform route fails this suite until the manifest is regenerated.

Release

Tag vX.Y.Z here, then run the publish-mcp workflow in omnia-tracing with that tag (dir: . publishes @error-bar/mcp; run it again with dir: legacy to publish the @omnia-voice/mcp forwarding shim) — the npm token lives there.

License: Apache-2.0