mcpField
P
Paw Pulse

Paw Pulse

Remote
pawpulse.dougalpaw.workers.devCloud & Infra0Updated 2d ago

About Paw Pulse

Paw Pulse is a push-based heartbeat and performance monitor for self-hosted LLM servers, aimed at operators of Ollama, llama.cpp, and vLLM. It detects downtime, slow generation, and Ollama GPU fallback, then sends webhook alerts.

Key features

  1. 01Send heartbeats and track availability
  2. 02Monitor generation-speed thresholds
  3. 03Detect Ollama GPU fallback
  4. 04Send state alerts by webhook
  5. 05Provide public status pages and badges

Use cases

  1. 01Detect an offline home LLM host
  2. 02Investigate slower Ollama generation
  3. 03Monitor a NAS, backup, or cron task

MCP support

The remote MCP server exposes tools to create, configure, inspect, and delete monitors. Connecting requires no authentication; monitor-changing tools require a manage_token.

Configuration

Remote servers need no local install. Pick your client, then copy the config or install in one click.

Endpointhttps://pawpulse.dougalpaw.workers.dev/mcp
Copy install command
Claude CodeTerminal
Cursor.cursor/mcp.json
VS Code.vscode/mcp.json
Claude Desktopclaude_desktop_config.json
Windsurf~/.codeium/windsurf/mcp_config.json

Run commands in a terminal; paste JSON into the client’s MCP config and restart. Servers that need sign-in will prompt on first connect.

Tools

6
ParameterTypeRequiredDescription
namestringNoDisplay name, not treated as identifying.
publicbooleanNoWhether /v1/monitors/:id/status and /s/:id are public.
webhook_urlstringNoOptional https:// webhook for alerts.
grace_secondsnumberNo60-86400, default 300.
slow_tps_thresholdnumberNoOptional tokens/sec floor.

Details

From the project README

Paw Pulse

Push-based heartbeat and performance monitor for self-hosted LLM servers (Ollama, llama.cpp, vLLM, anything OpenAI-compatible). Know when your home AI box dies or slows down, without exposing it to the internet.

Live instance: https://pulse.dougalpaw.com

Demo: Ollama fails, Paw Pulse posts a DOWN alert to Discord, then a recovery

Built, deployed, and operated by Pup, an AI agent, as part of the Dougal Paw project. No accounts, no IP storage, no personal data: monitors are anonymous, identified only by a random id plus an ingest token and a manage token.

How it works

A tiny client script (pawpulse.sh) runs on the box hosting your LLM server and POSTs a heartbeat out to this service every few minutes. If heartbeats stop arriving within the configured grace period, the monitor flips to down. Your LLM server never accepts an inbound connection.

No LLM on the box? LLM_KIND=none sends a plain heartbeat, so the same alerts work for a NAS, a backup job or any cron task.

On Ollama it can also catch the quiet failures: set slow_tps_threshold to hear when generation crawls, or min_gpu_pct to hear when a model has fallen back from GPU to CPU (the client reads size_vram from /api/ps; needs jq).

Guides:

Stack

  • Cloudflare Workers (free tier) + D1 + Cron Triggers.
  • TypeScript, Hono as a thin router.
  • No other runtime dependencies.

Project layout

src/
  index.ts       Hono app: routes, pages, cron entrypoint
  service.ts     Business logic shared by REST and MCP
  db.ts          D1 queries
  state.ts       Pure state-machine logic (unit tested)
  validation.ts  Pure input validation (unit tested)
  cron.ts        Per-minute sweep: state transitions + alert dispatch
  webhook.ts     Outbound alert delivery (Discord-aware, generic JSON fallback)
  mcp.ts         Remote MCP server (Streamable HTTP, stateless)
  openapi.ts     OpenAPI 3.1 document served at /openapi.json
  llms.ts        /llms.txt content
  templates.ts   HTML pages (/, /privacy, /s/:id)
  badge.ts       Shields-style SVG badge renderer
  static/        Served static assets (pawpulse.sh)
migrations/      D1 schema (0001_init.sql)
test/            node:test unit tests for state.ts and validation.ts

API

See /openapi.json for the full spec and /llms.txt for a terse summary.

  • POST /v1/monitors — create a monitor.
  • POST /v1/monitors/:id/beat — send a heartbeat (Authorization: Bearer <ingest_token>).
  • GET /v1/monitors/:id/status — current state; public monitors need no auth.
  • GET /v1/monitors/:id/heartbeats — recent heartbeats.
  • PATCH / DELETE /v1/monitors/:id — manage/remove (Authorization: Bearer <manage_token>).
  • GET /s/:id — public HTML status page. GET /badge/:id.svg — status badge.
  • GET /pawpulse.sh — the client script.
  • POST /mcp — remote MCP server (see below).

MCP

https://pulse.dougalpaw.com/mcp is a remote MCP server (Streamable HTTP). It is stateless: each POST carries one JSON-RPC message and gets a JSON response, with no sessions and no SSE stream, which the spec allows. It supports protocol versions 2025-06-18, 2025-03-26 and 2024-11-05.

Tools: create_monitor, setup_instructions, get_status, list_heartbeats, update_monitor, delete_monitor. They use the same service layer as the REST API. No auth is needed to connect; tools that change a monitor take its manage_token as an argument.

{ "mcpServers": { "pawpulse": { "type": "http", "url": "https://pulse.dougalpaw.com/mcp" } } }

Alerts

Set a monitor's webhook_url and you get a message when it goes down, gets slow or recovers. The format is picked from the URL: Discord, Slack, Microsoft Teams (Workflows), Telegram (https://api.telegram.org/bot<token>/sendMessage?chat_id=<id>) and Google Chat get native messages; any other https URL gets generic JSON (content, text, monitor_name, alert, state, timestamp). Self-hosters can switch all alerts off with ALERTS_ENABLED = "false" in wrangler.toml.

Development

pnpm install
npm test         # unit tests (node:test, no network/D1 needed)
npm run typecheck

Deploy

npx wrangler d1 create pawpulse                 # once; copy the database_id into wrangler.toml
npx wrangler d1 migrations apply pawpulse --remote
npx wrangler deploy

Cron runs every minute (* * * * *) and sweeps monitors for grace-period expiry and slow-tps thresholds, firing at most one webhook per state change and capping outbound webhook fetches per run to stay under the Workers free-tier 50-subrequest limit.

Privacy

See /privacy. No accounts, no cookies, no analytics, no IP logging — anywhere. DELETE /v1/monitors/:id permanently removes a monitor and all its heartbeats.

Licence

MIT. See LICENSE.

Comments

Markdown supported

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.