TracePulse

Runtime feedback MCP server for AI coding agents.

Fewer wasted tokens. Faster shipping. Lower carbon footprint. Responsible AI in action.

ViewGraph sees the UI. TracePulse hears the backend.

"LLMs can't see what happens when their code actually runs. They're throwing darts in the dark." - Sentry Engineering

TracePulse closes this loop at dev time - seconds after the code change, not minutes after deployment.

TracePulse watches your dev server's stdout/stderr, parses errors into structured events with signal scoring, and exposes them as MCP tools that any AI coding agent can call. The agent edits code, calls get_errors, and instantly knows if the fix worked - no manual log reading, no copy-paste.

Status

🟡 Alpha v0.9.14 - Phases 1-5 complete. Core pipeline, watch mode, multi-process support, frontend-backend correlation, proactive monitoring. 26 error parsers, 39 MCP tools, 986 tests passing.

Quick Start

Add TracePulse to your MCP client's config file. The file location depends on which tool you use:

Config File Locations

⚠️ Common mistake: Kiro CLI uses .kiro/settings/mcp.json, not .kiro/mcp.json. If TracePulse tools don't appear, check you're editing the right file.

Zero-config (works for any project)

{
  "mcpServers": {
    "tracepulse": {
      "command": "tracepulse"
    }
  }
}

TracePulse auto-detects your project type (Node, Python, Go, Rust, Java) and provides tools immediately. No server command needed - the agent can call start_server() later when ready.

For non-Node projects, install globally first: npm install -g tracepulse

Start mode - spawn and monitor your dev server

Node.js projects (npm/pnpm/Bun available in project):

{
  "mcpServers": {
    "tracepulse": {
      "command": "npx",
      "args": ["tracepulse", "start", "npm run dev"]
    }
  }
}

Python, Go, Java, Rust, or any non-Node project:

Node.js may not be on your project's PATH. Install TracePulse globally once:

npm install -g tracepulse

Then use the global binary directly:

{
  "mcpServers": {
    "tracepulse": {
      "command": "tracepulse",
      "args": ["start", "python manage.py runserver"]
    }
  }
}

Replace the command with your dev server: uvicorn main:app --reload, go run main.go, mvn spring-boot:run, cargo run, etc.

If your server needs environment variables (e.g., PYTHONPATH), use the env field:

{
  "mcpServers": {
    "tracepulse": {
      "command": "tracepulse",
      "args": ["start", "python -m myapp.server"],
      "env": { "PYTHONPATH": "src" }
    }
  }
}

If your server uses a start script (bash, shell):

{
  "mcpServers": {
    "tracepulse": {
      "command": "tracepulse",
      "args": ["start", "bash scripts/start.sh"]
    }
  }
}

⚠️ Environment variables go in the env field, not in the command. "args": ["start", "PYTHONPATH=src python app.py"] won't work. Put variables in env and keep the command clean.

Attach mode - tail an existing log file

Use this when your servers are already running (managed by scripts, Docker, process managers, etc.):

{
  "mcpServers": {
    "tracepulse": {
      "command": "npx",
      "args": ["tracepulse", "attach", "--log-file", "./logs/server.log"]
    }
  }
}

Local development (without npm publish)

Point directly to the built CLI:

{
  "mcpServers": {
    "tracepulse": {
      "command": "node",
      "args": ["/path/to/tracepulse/dist/cli.js", "start", "npm run dev"]
    }
  }
}

Which mode should I use?

MCP Tools

Core (Phase 1)

Watch Mode (Phase 2)

Multi-Process (Phase 3)

Correlation (Phase 4)

Proactive (Phase 5)

Execution & Health

Error Intelligence

Error Parsers

TracePulse parses errors from 25 sources out of the box:

Runtime errors:

• Node.js - TypeError, ReferenceError, SyntaxError, etc. with V8 stack traces
• Python - Tracebacks with file:line extraction
• Go - Panics with goroutine stack traces
• Java - Exceptions with at frames and Caused by: chains
• Rust - Panics with RUST_BACKTRACE output
• JSON - Structured logs (pino, structlog JSON, logback) with level/message fields
• Structlog - Python structlog key-value format ([info], [warning], [error] brackets)

Build errors:

• TypeScript - tsc compiler errors (TS####)
• ESLint - Lint errors with rule names
• Vite/webpack - Build tool errors (module not found, transform failures)
• Build Stats - Module count, build time from Vite/webpack

Test runners:

• pytest - FAILED, ERROR, summary lines
• Jest - FAIL header, assertion details
• vitest - FAIL file, Expected/Received
• Go test - --- FAIL, error with file:line
• cargo test - test FAILED, panic with file:line, summary
• JUnit/Maven/Gradle - Surefire summary, Gradle task FAILED, AssertionError

Infrastructure:

• HTTP Access Log - uvicorn, express/morgan, nginx with status and duration
• Migration - alembic and Django migration output

Background workers:

• Celery - task raised/retry/timeout/succeeded
• Sidekiq - WARN/ERROR/FATAL job events, done timing
• BullMQ - job failed/stalled/completed, queue errors

Signal Scoring

Every event gets a signal_score (0–100) and signal_strength (high/medium/low):

CLI Usage

# Single process - spawn and monitor
tracepulse start "npm run dev"

Spawns your dev server as a child process, captures its stdout/stderr, and monitors for errors. TracePulse manages the process lifecycle - forwards SIGTERM/SIGKILL on shutdown.

# Attach to existing log file
tracepulse attach --log-file ./server.log

Tails an existing log file without spawning any process. Use when your servers are already running - managed by Docker, tmux, pm2, systemd, or custom scripts.

# Multi-process - monitor multiple services
tracepulse start --service api="npm run dev:api" --service worker="npm run worker"

Spawns multiple services simultaneously. Each service's output is tagged with its name - filter with get_errors(service: "api") or see all with list_services().

# Config file
tracepulse start --config tracepulse.config.json

Reads service definitions, transport settings, and persistence options from a JSON config file instead of CLI flags. See Configuration for the schema.

# Docker Compose
tracepulse compose --file docker-compose.yml

Discovers services from a Docker Compose file and tails their container logs via the Docker Engine API. Each container's output is tagged with its compose service name.

# With persistence (saves fingerprints across sessions)
tracepulse start --persist "npm run dev"

# Clustered mode (7 gateways instead of 39 tools - 80% schema reduction)
tracepulse start --clustered "npm run dev"

Collapses 39 tools into 7 semantic gateways. The agent discovers sub-tools on demand, loading only what it needs. Schema overhead drops from ~1,000 to ~200 tokens per turn.

# With HTTP transport (for multi-client scenarios)
tracepulse start --http "npm run dev"

Starts a Streamable HTTP server on 127.0.0.1:9800 alongside the default stdio transport. Allows multiple MCP clients to connect to the same TracePulse instance simultaneously.

Why run_and_watch Instead of Shell

run_and_watch isn't just a convenience - it solves real reliability problems:

Structured output. Shell commands return raw text the agent must parse. run_and_watch returns structured JSON with pass/fail counts, error details, and file:line references.

WSL reliability. On WSL (Windows Subsystem for Linux), terminal output capture is unreliable - Kiro IDE and other tools often can't read test results from the terminal. run_and_watch bypasses this entirely because it captures output via Node.js pipes and returns data over the MCP protocol (JSON-RPC over stdio), a completely separate channel from the terminal.

Monorepo support. The cwd parameter runs commands in subdirectories without cd prefix hacks:

run_and_watch("npx vitest run", cwd: "./frontend")
run_and_watch("pytest tests/", cwd: "./backend")

Parser pipeline. Output flows through TracePulse's 25 parsers, so test failures, build errors, and runtime crashes are all normalized into the same structured format the agent already knows.

Cloud Log Monitoring

Monitor cloud service logs with zero additional dependencies - uses your existing cloud CLIs:

The same 26 parsers that catch local dev server errors catch cloud errors too.

Hot-Reload Detection

TracePulse detects hot-reload events from 8 dev tools:

• Vite - compilation success, HMR updates
• webpack - compilation completed
• nodemon - restart/starting events
• Next.js - compilation, route compiling
• ts-node-dev - restart, compilation complete
• uvicorn - file change detection, reloader process
• Django - file change watching, system checks
• Flask - restart with stat/watchdog, change detection

When watch_for_errors detects a hot-reload, it sets hot_reload_detected: true in the response so the agent knows the server actually reloaded.

Security

• All log output is redacted for secrets (16 patterns: API keys, Bearer/JWT tokens, connection strings, PEM keys, GitHub/GitLab/Slack tokens, GCP service accounts, Azure connection strings, Datadog keys) before entering the pipeline
• No secrets in MCP responses, ring buffer, or diagnostic output
• HTTP endpoints bind to 127.0.0.1 only - no external access
• Fingerprint persistence stores only hashes, not raw error messages
• No file system writes except optional fingerprint persistence

TracePulse for Teams (Coming Soon)

Local: Your agent sees your server. Team: Every agent sees every server.

Deploy a shared TracePulse instance for your engineering team. Every developer's AI agent connects via HTTPS to one server that:

• Aggregates errors across all dev environments
• Shares fingerprints - if one developer hits a bug, every agent knows
• Team audit trail - see tool usage and token savings across the team
• Centralized drift detection - one health check covers shared staging
• SSO/API key auth - enterprise-ready access control

{
  "mcpServers": {
    "tracepulse": {
      "type": "streamable-http",
      "url": "https://tracepulse.internal.company.com/mcp",
      "headers": { "Authorization": "Bearer <team-api-key>" }
    }
  }
}

Self-hosted via Docker or one-click deploy on Railway/Fly.io. Roadmap

Companion Tools

TracePulse is designed to work alongside:

• ViewGraph - Visual cortex: structured UI perception (DOM, a11y, layout, annotations)
• Chrome DevTools MCP - Motor cortex: browser interaction (console, network, performance)

TracePulse is the auditory cortex - it hears what the backend is saying. Together, the agent has the same situational awareness a senior developer has: see the UI, act in the browser, hear the server.

The cortex analogy maps to the Perception-Planning-Acting modular architecture formalized in GUI agent research (arXiv 2412.13501, arXiv 2504.20464). ViewGraph provides the perception module, Chrome DevTools MCP the action module, TracePulse the environment feedback module.

Architecture

See docs/architecture/architecture-guide.md for the full architecture guide with diagrams.

License

AGPL-3.0