MCP Proxy

Observability and monitoring layer for AI agent tool calls via the Model Context Protocol (MCP). Know exactly what your AI agents are doing, how often, and when something goes wrong.

The problem

AI agents work autonomously — they call external APIs, perform actions, and make decisions. But developers have no visibility into what the agent is actually doing, how frequently, or when something breaks.

HookWatch MCP Proxy gives you full observability for MCP tool calls — the same way HookWatch already does it for webhooks and cron jobs.

How it works

HookWatch MCP Proxy sits between your AI agents and MCP servers. It transparently proxies all JSON-RPC 2.0 requests while collecting full telemetry — requests, responses, latency, errors, and tool usage.

Architecture

1
Register your MCP server

Add your MCP server via the dashboard or API. You'll get a unique proxy URL like https://mcp.hookwatch.dev/p/x7k9m2.
2
Point your agent to the proxy

Replace the original MCP server URL with your HookWatch proxy URL in your agent configuration (Claude Desktop, Cursor, or any custom agent). No code changes required.
3
Monitor everything

Every tool call is logged with full request/response bodies, latency, status, and tool name. View everything in your dashboard or query via API.

Quick start

1. Create an MCP server

Go to MCP Servers in your dashboard and click "Create Server", or use the API:

Create Server

Response:

Response

The slug field is your unique proxy identifier. Your proxy URL is https://mcp.hookwatch.dev/p/slug.

2. Configure your AI agent

Replace the MCP server URL in your agent configuration with the proxy URL.

Claude Desktop / Claude Code:

claude_desktop_config.json

Cursor:

.cursor/mcp.json

Any HTTP client / custom agent:

Shell

3. Monitor in the dashboard

Open MCP Calls in your dashboard to see every tool call in real time — method, tool name, status, duration, and full request/response bodies.

What gets logged

Every JSON-RPC request that passes through the proxy is logged with full context:

MCP Call Object

Field	Type	Description
`method`	string	JSON-RPC method (`tools/call`, `tools/list`, `initialize`, etc.)
`tool_name`	string	Extracted tool name from `tools/call` params (e.g., `get_weather`)
`request_body`	string	Full JSON-RPC request body
`response_body`	string	Full JSON-RPC response body from upstream
`status`	string	Call status: `success`, `error`, `timeout`
`status_code`	int	HTTP status code from the upstream MCP server
`duration_ms`	int	Total round-trip time in milliseconds
`source_ip`	string	IP address of the calling AI agent
`headers`	object	HTTP request headers from the agent

Proxy endpoint

The proxy accepts JSON-RPC 2.0 requests at:

POST https://mcp.hookwatch.dev/p/:slug

Example: Call a tool

Shell

Proxy behavior

Transparent forwarding

Requests are forwarded to your MCP server exactly as received. The proxy adds no modifications to the JSON-RPC payload.

Header passthrough

Content-Type, Accept, and Authorization headers are forwarded to the upstream server.

Async logging

Call logging is done asynchronously after the response is returned. This ensures minimal latency overhead on the proxy path.

Circuit breaker

Built-in circuit breaker protects against cascading failures. If the proxy's internal services are down, it returns 503 instead of hanging.

Error detection

The proxy detects both HTTP errors (status 400+) and JSON-RPC errors in the response body, marking calls with the correct status.

Limits

60-second upstream timeout, 10MB max payload size. Requests to inactive servers return 404.

Features

Full request/response logging

Every tool call is logged with complete request and response bodies, parameters, errors, and timing data. Debug agent behavior in seconds.

Tool-level tracking

The proxy automatically extracts tool names from tools/call requests. Filter and search your call history by specific tool.

Latency monitoring

Track MCP server response times per call and through aggregated metrics. Identify slow tools and optimize performance.

Dashboard

Full-featured UI for managing MCP servers, browsing call history with filters, and viewing detailed request/response data for each call.

Alerts

Set up notifications for errors, timeouts, and anomalous activity via email, Telegram, and webhooks. Know immediately when something goes wrong.

Prometheus metrics

Built-in Prometheus endpoint exposes call latency histograms and call counts by slug, method, and status for integration with your monitoring stack.

API reference

Manage MCP servers and query call logs through the REST API. All endpoints require authentication via Bearer token.

MCP Servers

Base URL

https://api.hookwatch.dev/v1/mcp/servers

POST /v1/mcp/servers — Create a new MCP server

GET /v1/mcp/servers — List all MCP servers for the team

GET /v1/mcp/servers/:id — Get a specific MCP server

PUT /v1/mcp/servers/:id — Update name, target URL, or active status

DELETE /v1/mcp/servers/:id — Delete server and all associated call logs

See MCP Servers API for full request/response details and examples.

MCP Calls

Base URL

https://api.hookwatch.dev/v1/mcp/calls

GET /v1/mcp/calls — List calls with filters (server, method, tool, status, date range)

GET /v1/mcp/calls/:id — Get full call details with request/response bodies

GET /v1/mcp/calls/count — Count calls matching filters

GET /v1/mcp/calls/metrics — Aggregated metrics (configurable time range)

DELETE /v1/mcp/calls/:id — Delete a specific call log

Query filters

Parameter	Type	Description
`server_id`	string	Filter by MCP server ID
`method`	string	Filter by JSON-RPC method (e.g., `tools/call`)
`tool_name`	string	Filter by tool name
`status`	string	Filter by status: `success`, `error`, `timeout`
`from` / `to`	string	Date range filter (ISO 8601)
`limit` / `offset`	int	Pagination (default limit: 50)

Example: List error calls

Shell

Example: Get metrics

Shell

See MCP Calls API for full response schemas and more examples.

Use cases

Monitor autonomous agents

Your AI agent runs 24/7 calling external APIs. HookWatch Proxy shows every call, response time, and error — you always know what the agent is doing.

Debug tool failures

Agent getting unexpected responses from an MCP server? Full request/response logs with timing data help you find the problem in seconds.

Compliance & audit

For regulated industries — complete audit trail of all AI agent actions. Every tool call logged with who, what, when, and full parameters.

Cost optimization

Track tool call frequency and patterns. Identify inefficient usage, redundant calls, and optimize your API spend.

Plans & limits

The number of MCP servers you can register depends on your plan:

Plan	MCP Servers
Free	Not available
Starter	3 servers
Pro	10 servers
Team	Unlimited

Why HookWatch for MCP?

Advantage	Details
Proven infrastructure	Alerting, logging, and dashboards already battle-tested for webhooks and cron jobs
Unified tooling	Monitor webhooks + cron jobs + AI agent tools in one place
Zero-code integration	Replace one URL — no changes to agent code or MCP server
High performance	Go-based reverse proxy with async logging — minimal latency overhead
Resilient	Circuit breaker, graceful degradation, and health monitoring built in

JSON-RPC 2.0

HookWatch MCP Proxy works with any MCP server that speaks JSON-RPC 2.0 over HTTP. All standard MCP methods are supported: initialize, tools/list, tools/call, resources/list, resources/read, and more.

MCP Proxy

The problem

How it works

Register your MCP server

Point your agent to the proxy

Monitor everything

Quick start

1. Create an MCP server

2. Configure your AI agent

3. Monitor in the dashboard

What gets logged

Proxy endpoint

Example: Call a tool

Proxy behavior

Transparent forwarding

Header passthrough

Async logging

Circuit breaker

Error detection

Limits

Features

Full request/response logging

Tool-level tracking

Latency monitoring

Dashboard

Alerts

Prometheus metrics

API reference

MCP Servers

MCP Calls

Query filters

Example: List error calls

Example: Get metrics

Use cases

Monitor autonomous agents

Debug tool failures

Compliance & audit

Cost optimization

Plans & limits

Why HookWatch for MCP?

Related