Nats Trail

Give an AI agent safe access to your production event bus.
_{A web UI, a CLI and an MCP server over one bounded query engine — so humans and agents debug NATS from the same source of truth.}

Why this exists

Every NATS GUI answers "what is in this stream?". That is the easy question.

The hard question in an event-driven system is "why did this one flow fail?" — and answering it means following a single request_id across four streams, three services and a dead-letter subject. Today you do that by hand, with a terminal per stream.

NATS Trail answers that question directly, and exposes the answer to agents as a typed, bounded, read-only tool contract — so you can point Claude at production and ask.

nats-trail trace --request-id req-0001d --limit 20

Or, from an agent: "Why did the refresh for s3-events fail?" — and it lands on the message that actually broke, with the retry state and the correlation ids already extracted:

The part nobody else does

There are several NATS MCP servers. They shell out to the nats CLI, return raw dumps, and expose publish and delete while describing themselves as read-only.

NATS Trail treats the agent surface as a contract, not a wrapper:

"Read-only" here means unreachable, not disabled

NATS Trail can write: publish, purge, delete messages, consumers and streams. Those live behind /api/mutate, reachable from the UI and the CLI.

executeMcpTool() receives an McpRuntimeData interface that exposes only read functions. There is no disabled publish behind a feature flag — there is no publish to call. A misconfigured environment variable cannot purge your production stream, because the code path does not exist. CLI write commands additionally refuse to run under --agent, and the test suite reads the source to assert none of this has been quietly undone.

This is the only reason it is reasonable to hand an agent a prod context.

Install

npx nats-trail serve

Open http://127.0.0.1:4000 — one process serves the UI and the API.

Prefer containers? The compose file brings up NATS Trail next to a JetStream-enabled server:

docker compose up

git clone https://github.com/solsolettidev/nats-trail
cd nats-trail
npm install
npm start

npm run dev

Requirements: Node.js >= 22 (for the built-in SQLite used by the correlation index) and a reachable NATS server (nats-server -js is fine).

Use it as an MCP server

Point any MCP client at the stdio server. For Claude Code:

claude mcp add nats-trail -- npx -y @nats-trail/mcp

Or wire it manually:

{
  "mcpServers": {
    "nats-trail": {
      "command": "natstrail-mcp",
      "env": {
        "NATS_TRAIL_API": "http://127.0.0.1:4000",
        "NATS_TRAIL_TOKEN": "<bearer token>"
      }
    }
  }
}

Twenty-five read-only tools, all returning the same envelope:

# start here when the topology is unknown
natstrail.discover_subjects        subjects that carry traffic + inferred payload shapes
natstrail.get_health_summary       what is broken right now, ranked
natstrail.reconstruct_flow         the causal chain for one request_id
natstrail.enrich_incident          flat context for an incident: flow + dead letters + health

# messages
natstrail.search_messages          natstrail.get_message_detail
natstrail.trace_by_request_id      natstrail.trace_by_correlation_id
natstrail.search_dlq               natstrail.run_filter

# topology
natstrail.list_streams             natstrail.get_stream_info
natstrail.list_consumers           natstrail.list_kv_buckets
natstrail.list_kv_keys             natstrail.get_kv_history
natstrail.list_object_buckets      natstrail.list_objects

# operations
natstrail.get_server_health        natstrail.list_server_connections
natstrail.list_contexts            natstrail.get_connection_status
natstrail.list_filters             natstrail.list_audit
natstrail.enrich_sentry

See docs/mcp-agent.md.

Three surfaces, one engine

                 ┌──  Web UI          explore visually, save filters
Query Engine  ───┼──  CLI             scripts, pipelines, humans in a terminal
  (core)         └──  MCP + HTTP API  agents, Sentry, dashboards

A filter you save in the UI is the same filter nats-trail filter run executes, and the same one natstrail.run_filter hands an agent. Configure a context once; use it everywhere.

• packages/ui — React + Vite. Presentation only.
• packages/server — Express + WebSocket bridge. Owns connections and credentials.
• packages/core — the Query Engine: envelopes, limits, truncation, filters, error normalization.
• packages/cli — the nats-trail binary: query commands, serve, and an interactive shell.
• packages/mcp — tool contracts and the stdio server.

See docs/architecture.md.

Features

Inspect — context selector (local / dev / staging / prod) with a prod confirmation gate, live subject subscription, JSON pretty print with tree view and in-payload search, stream and consumer browsing, auto-detected DLQ panel.

JetStream — streams, subjects, counts and consumer health.	NATS Core — live subject subscription with filters.
DLQ — auto-detected dead letters with reason and origin.	Viewer — tree, raw, search, fullscreen, copy-path.

Understand — subject discovery infers payload shapes from real traffic, flow reconstruction turns one request_id into a causal chain, and the health summary ranks what is actually broken.

Query — bounded stream scans with cursors and time windows, filters by subject, date, text and JSON event type, saved filters shared across all three surfaces. Binary payloads (protobuf, msgpack) are detected and shown as hex dumps rather than mojibake.

Browse — KV buckets with per-key revision history including deletes, Object Store metadata, and server health from the monitoring port (varz / jsz / connz).

Change — publish, request/reply, purge, and delete messages, consumers and streams, from the UI and the CLI, behind confirmation. Never from the agent surface.

Integrate — read-only HTTP API under /api/integration, bearer tokens with per-token audit identity, and POST /api/integration/enrich/sentry to attach NATS context to an error without exposing credentials.

Full list in docs/features.md · CLI reference in docs/cli.md.

Security