@tracelanedev/mcp

@tracelanedev/mcp — Tracelane MCP Server

Read-only MCP server exposing Tracelane trace data to any MCP-compatible client — Claude Desktop, Claude Code, Cursor, or any agent using the Model Context Protocol.

Not on npm yet — npx @tracelanedev/mcp still 404s, including in the config blocks below. The package is wired into the release workflow and becomes installable when the next signed release tag carries it. Releases are bundled — one tag covering everything that has moved, never a tag cut for a single package (VERSIONING.md, "Release cadence") — so treat every npx line here as the post-publish form and use the from-source path today. The same run submits apps/mcp/server.json to the MCP registry; the name it will be listed under is io.github.tracelane/tracelane-mcp.

Default mode reads through the gateway — the same tenant-scoped /v1/* routes the dashboard uses — with just TRACELANE_API_KEY and TRACELANE_GATEWAY_URL. That is the Cloud-tenant path (PLT-22) and needs no ClickHouse credentials. Set CLICKHOUSE_URL to switch to self-host mode, reading ClickHouse directly instead — see Self-hosting.

Quick start

Cloud (Tracelane-hosted tenant)

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS), %APPDATA%\Claude\claude_desktop_config.json (Windows), or .mcp.json at your project root for Claude Code:

{
  "mcpServers": {
    "tracelane": {
      "command": "npx",
      "args": ["@tracelanedev/mcp"],
      "env": {
        "TRACELANE_API_KEY": "tlane_YOUR_KEY",
        "TRACELANE_GATEWAY_URL": "https://gateway.tracelane.dev"
      }
    }
  }
}

No CLICKHOUSE_URL — its absence is what selects gateway mode. Every read goes through the gateway's existing tenant-scoped routes, so a key from another tenant, or one lacking the read scope, gets a clear tool error rather than an empty result.

Until it is on npm

Swap the two launch keys for a path into your clone — every env key is unchanged:

"command": "node",
"args": ["/path/to/tracelane/apps/mcp/dist/index.js"]

Build it first with pnpm install && pnpm --filter @tracelanedev/mcp build.

The Streamable HTTP transport ships in this package (TRACELANE_MCP_TRANSPORT=http, see Transports) — run it yourself. There is no hosted endpoint: https://mcp.tracelane.dev does not resolve, so a url-style client entry has nothing to connect to.

Self-host (ClickHouse)

Set CLICKHOUSE_URL to read ClickHouse directly instead of the gateway — for a self-hosted or local Tracelane stack:

{
  "mcpServers": {
    "tracelane": {
      "command": "npx",
      "args": ["@tracelanedev/mcp"],
      "env": {
        "TRACELANE_API_KEY": "tlane_YOUR_KEY",
        "TRACELANE_GATEWAY_URL": "https://gateway.tracelane.dev",
        "CLICKHOUSE_URL": "http://localhost:8123",
        "CLICKHOUSE_USER": "default",
        "CLICKHOUSE_PASSWORD": "…"
      }
    }
  }
}

TRACELANE_GATEWAY_URL is still needed in self-host mode — it is where TRACELANE_API_KEY is validated (/v1/auth/whoami) even though trace reads go to ClickHouse instead.

Tools

Tool	Description
list_traces	List recent traces for the tenant. Params: limit (default 20), model_filter, has_error
get_trace	Get all spans for a trace. Params: trace_id
get_span	Get full details for a span including all LLM GenAI attributes. Params: span_id, trace_id — required in Cloud (gateway) mode (no span-by-id gateway route; the span is picked out of the trace's span list), optional in self-host mode
search_traces	Free-text search across span names and attributes. Params: query (gateway mode requires ≥4 chars), model_filter?, has_error?, limit. Gateway mode returns content-filtered trace summaries; self-host mode returns per-span match detail (matched_spans, first_match_*)
explain_guardrail_block	Human-readable explanation of a guardrail signal. Self-host mode: trace_id + span_id (a detection-layer AFT flag on a recorded span). Gateway mode: EITHER correlation_id (from a block's 403 body, for a request blocked pre-flight with no trace) OR trace_id + span_id (same AFT-flag case, still available since spans carry the flag either way)
list_evals	List every pain-point + fault-tolerance eval id and count, read from the manifest bundled at build time from evals/. Params: none
get_eval_result	Read a specific eval's assertions. Needs a repo checkout for the source; says so when there is none. Params: eval_id
replay_trace	Return a recorded trace as-is (ordered spans with LLM/tool attributes) for offline step-through. Read-only — it does not re-execute any model or tool. Params: trace_id, include_tool_calls?

Example usage in Claude

Once connected, you can ask Claude:

"Show me the last 5 traces that had a guardrail block, and explain what fired."

"Compare the latency of traces using claude-haiku-4-5 vs claude-sonnet-4-6 in the last hour."

"Show me the assertions that eval makes."

Auth

V1: TRACELANE_API_KEY environment variable passed via the MCP env block. The server resolves the tenant from the API key — tenant_id is never accepted as a tool argument.

V2 (roadmap): OAuth 2.1 PKCE. The authorization server is https://gateway.tracelane.dev/.well-known/oauth-authorization-server. tenant_id extracted from JWT organizationId claim only.

Transports

Transport	How to select it	When to use
Stdio	default	Local use — Claude Desktop, Claude Code, Cursor. Zero network exposure.
Streamable HTTP	TRACELANE_MCP_TRANSPORT=http TRACELANE_MCP_PORT=8081	Self-run remote deployments. Every request must carry Authorization: Bearer <jwt-or-tlane-key>; the tenant is resolved per request through the gateway. No hosted endpoint is operated for you.

Security invariants

• Read-only. No write tools are registered — the tool surface is the eight listed above, all of which only read.
• Tenant isolation. Gateway mode: every read is a tenant-scoped gateway route (crates/gateway/src/trace_reads.rs) — the tenant comes from the bearer's claims server-side, this server never sends or sees a tenant id. Self-host mode: every ClickHouse query includes WHERE tenant_id = {tenantId: String} (parameter-bound, never string-interpolated).
• tenant_id is never a tool parameter in either mode. Stdio resolves it once at startup from TRACELANE_API_KEY via the gateway and refuses to start if the key is rejected; HTTP resolves it per request from the bearer token and binds it through AsyncLocalStorage.
• A non-2xx gateway response is a tool error, never an empty result. A revoked or wrong-tenant key, a key missing the read scope, or another tenant's trace id all read back as isError: true carrying the gateway's own status and message — not [].
• No eval id reaches the filesystem. get_eval_result looks the id up in the bundled manifest and uses the manifest's path, so a traversal string cannot name a file.
• TRACELANE_GATEWAY_URL is SSRF-checked before any bearer is sent to it, in both modes: https-only outside development, tracelane.dev hosts only, private/CGNAT/IMDS ranges refused.

Known gap — span content is returned verbatim. There is no redaction pass over span attributes and no untrusted-content sentinel around user text. Do not point this server at a workspace whose spans may carry secrets you would not hand to the connected model.

Self-hosting

# From source, against your own stack
pnpm dev:mcp

No container image is published for the MCP server — ghcr.io/tracelane/mcp does not exist. Run it from source or from the npm package once it ships.

Stack

• @modelcontextprotocol/sdk — official MCP SDK (stdio + Streamable HTTP)
• @clickhouse/client — parameter-bound ClickHouse queries
• TypeScript 5.5 strict, noUncheckedIndexedAccess: true
• Biome for lint + format

License

Apache 2.0 — see LICENSE.