About tracelane-mcp
tracelane-mcp is a read-only MCP server for inspecting Tracelane agent traces, evaluations, and guardrail verdicts. It serves teams analyzing agent behavior, with tenant-scoped access through the gateway or direct ClickHouse reads for self-hosted setups.
Key features
- 01List and search traces
- 02Inspect traces and spans
- 03Explain guardrail blocks
- 04List evaluations and read results
- 05Replay recorded traces without execution
- 06Gateway or ClickHouse access
Use cases
- 01Investigate traces with errors or guardrail blocks
- 02Inspect LLM and tool span details
- 03Review evaluation assertions
- 04Step through a recorded trace offline
MCP support
Exposes eight read-only trace and evaluation tools to MCP clients over stdio or self-run Streamable HTTP. Tenant-scoped reads use the Tracelane gateway by default, or ClickHouse when configured for self-hosting.
Configuration
Local server config is the same for every client. Copy it into your client’s MCP config file.
npx -y @tracelanedev/mcpTRACELANE_API_KEYSecretTracelane API key (tlane_…). Resolved to a tenant once at startup via the gateway's /v1/auth/whoami; the server refuses to start if the key is rejected. Also the credential used to read trace data through the gateway when CLICKHOUSE_URL is unset.TRACELANE_GATEWAY_URLBase URL of the Tracelane gateway. Validates TRACELANE_API_KEY, and — when CLICKHOUSE_URL is unset — is where every trace read goes. Must be https://…tracelane.dev outside development.CLICKHOUSE_URLself-host only; when unset the server reads through the gateway. Set this to point the server at your own ClickHouse instead of Tracelane Cloud.CLICKHOUSE_USERClickHouse username.CLICKHOUSE_PASSWORDSecretClickHouse password.CLICKHOUSE_DBClickHouse database holding the spans table.TRACELANE_MCP_TRANSPORTTransport to serve. stdio (default) for desktop clients; http for the Streamable HTTP transport.Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.
Tools
Local servers expose tools only once running on your machine, and the maintainer has not provided a tool list.
Details
From the project README@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.
On npm since 2026-09-07:
npx @tracelanedev/mcpinstalls the published version; the config blocks below work as written. From-source is still documented under self-hosting. The same run submitsapps/mcp/server.jsonto the MCP registry; the name it will be listed under isio.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 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.
Run from a clone
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
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.
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 includesWHERE tenant_id = {tenantId: String}(parameter-bound, never string-interpolated). tenant_idis never a tool parameter in either mode. Stdio resolves it once at startup fromTRACELANE_API_KEYvia the gateway and refuses to start if the key is rejected; HTTP resolves it per request from the bearer token and binds it throughAsyncLocalStorage.- A non-2xx gateway response is a tool error, never an empty result. A revoked or wrong-tenant key, a key missing the
readscope, or another tenant's trace id all read back asisError: truecarrying the gateway's own status and message — not[]. - No eval id reaches the filesystem.
get_eval_resultlooks the id up in the bundled manifest and uses the manifest's path, so a traversal string cannot name a file. TRACELANE_GATEWAY_URLis 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.
Operator note: span content is returned as stored. Scope the API key you give this server to workspaces whose spans you are comfortable handing 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.
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.
Comments
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.