Agent Watch MCP¶
Remote MCP server (Streamable HTTP) - continuous monitoring for the endpoints your agents depend on: liveness, latency, schema drift, auth posture and price integrity for MCP and paid agent services. Free tier covers 5 endpoints, and a free probe API checks any endpoint before you pay.
Server type: Remote (Streamable HTTP)
Auth: watchlist tied to checkout email; probe is open
Endpoint: https://agent-watch-api-production.up.railway.app/mcp/
Tools: watchlist and probe operations for MCP endpoint health
Pricing: Free 5 endpoints; Builder $19/mo (50); Team $99/mo (500 + API + alerts); Data API metered
Category: DevOps
Built by: entradox (sibling of Perimeter Watch and AgentLedger)
Why This Matters for Operators¶
Every production agent now leans on third-party MCP endpoints, and those endpoints fail silently: the vendor's own census found roughly two-thirds of registry-declared MCP endpoints do not answer, and an arXiv census found 47.9% of advertised x402 paid hosts dead. Agent Watch is the monitoring layer that directory listings never had - when an endpoint your agents pay for goes dark, drifts its schema, or starts charging differently than advertised, you get an alert instead of a failed workflow.
For buyers of agent services, the free probe is the due-diligence step before committing: POST any endpoint URL and get a liveness answer in seconds.
Tools & Capabilities¶
| Capability | What it checks |
|---|---|
| Liveness | MCP handshake + tool list |
| Latency | p50/p95 response time |
| Schema drift | Tool-list hash changes between probes |
| Auth posture | RFC 9728 metadata presence |
| Price integrity | 402 challenge vs advertised price (Phase 1) |
REST equivalents are documented: POST /v1/probe for free checks and POST /v1/watch to add endpoints.
Installation¶
claude mcp add agent-watch --transport http https://agent-watch-api-production.up.railway.app/mcp/
Configuration¶
{
"mcpServers": {
"agent-watch": {
"type": "http",
"url": "https://agent-watch-api-production.up.railway.app/mcp/"
}
}
}
Watchlist endpoints are tied to the email from checkout; the probe endpoint works without auth.
Business Relevance¶
- Operators running agent fleets get one dashboard for every MCP endpoint in the stack instead of discovering failures from users.
- Teams buying paid agent services verify liveness and price integrity before signing up, and keep verifying after.
- Platform builders get the metered Data API to fold endpoint health into their own routers and marketplaces.
- Security leads see schema drift - an unexpected tool-list change is an early signal of a compromised or silently-updated endpoint.
Integration with CorpusIQ¶
Agent Watch monitors the plumbing CorpusIQ runs on: as CorpusIQ's connector ecosystem grows, operators can watchlist the MCP endpoints behind critical automations and surface alerts next to CorpusIQ's own health dashboards. A composed workflow: CorpusIQ flags a connector error mid-workflow, Agent Watch confirms whether the underlying endpoint was alive at that moment - separating vendor outages from configuration bugs in one look.
Limitations¶
- Brand new - no track record yet; early-stage pricing and endpoint on a railway.app host.
- Free tier caps at 5 endpoints; real fleets need Builder or Team.
- Price-integrity monitoring is labeled Phase 1 - expect coverage gaps.
- No published self-host option; monitoring data lives on the vendor's endpoint.