MCPWatch wraps any Python MCP server with one line. Every tool call tracked. p95 latency, per-tool error rates, silent failures surfaced before your users find them. Open source, MIT, zero config required.
The problem
A tool returns empty results. Latency creeps from 200ms to 4 seconds. The agent works around it. You find out from a user three days later.
Some MCP errors don't bubble up as exceptions. Empty arrays. Wrong shape. The agent shrugs and moves on. Without instrumentation, the failure is invisible until a user complains.
Tool was 200ms last week, 4 seconds today. No per-tool baseline, no p95, no alert. You'll notice when the user complains, or when a third-party API quietly degrades.
If you log to stdout you can grep. If you don't, exceptions evaporate. Either way, no per-tool error rate, no anomaly detection, no idea which tool just got worse.
The wedge feature
One line of instrumentation. Every tool call tracked. Plus six more modules underneath when you need them.
One-line instrumentation for any Python MCP server. Works with FastMCP and the low-level mcp.server.lowlevel.Server. Tracks every tool call: latency, errors, silent failures (empty returns, isError responses). Per-tool p95, error rates, alerts.
Background process compiles system state into a structured AWARENESS.md every 90 seconds. Agents boot with full context — no cold starts.
Expose only the tools relevant to the current task. Switch frames with a query param. Eliminate 40–80K tokens of noise from every context window.
Lightweight structured signals between agents and systems. Post from anywhere — CLI, Python, HTTP. Queryable signal history. Coordination without queues.
Structured STATE.md written at session end. Next agent — or next session of the same agent — picks up exactly where you left off. No verbal summaries needed.
Vigil watches signals and session logs to extract durable knowledge — patterns, decisions, corrections. Stored and surfaced automatically, not buried.
Built-in MCP server exposes all Vigil tools to any MCP-compatible client. Works out-of-the-box with Claude, Cursor, Windsurf, and any MCP host.
How it works
No configuration files required to get started. Works with your existing MCP setup.
Single package from PyPI. No dependencies pulled in.
Creates a vigil.db and begins compiling AWARENESS.md every 90 seconds.
Point Claude or Cursor at the SSE URL. Add a frame param to scope tools.
Use cases
Vigil works anywhere agents need shared context, structured handoff, or scoped tool access.
Three agents review the same PR — security, style, and correctness. Each emits signals as they work. The next agent picks up where the last left off via structured handoff.
GitHub Actions emit signals on every deploy. Vigil triggers fire Slack webhooks on failure, add focus items on rollback, and keep your agents aware of production state.
Frame-based filtering gives each agent only the tools it needs. A research agent sees search tools; a writing agent sees document tools. Same MCP server, different frames — 50-90% fewer tokens.
Pricing
The open-source core is always free. Hosted cloud tier is coming Q3 2026.
Free
forever · open source
Pro
cloud hosted · coming Q3
Team
cloud hosted · coming Q3
Enterprise
cloud hosted · coming Q3
How it fits
Vigil is complementary to memory tools. It handles coordination, awareness, and tool scope — the layers that memory libraries don't cover.
| Capability | Vigil | Mem0 | Letta | LangGraph |
|---|---|---|---|---|
| Session awareness / context compile | ✓ | — | partial | — |
| Frame-based tool filtering | ✓ | — | — | — |
| Signal / event protocol | ✓ | — | — | partial |
| Structured session handoff | ✓ | — | partial | partial |
| Vector / semantic memory | — | ✓ | ✓ | — |
| Multi-agent orchestration | signals | — | ✓ | ✓ |
| Native MCP server | ✓ | — | — | — |
| Zero dependencies | ✓ | — | — | — |
Use Vigil alongside Mem0 or Letta — they operate at different layers. Vigil handles coordination and awareness; memory tools handle retrieval.
One line of instrumentation. Catch silent failures, latency spikes, and per-tool errors before they reach production.