mcp-observability

# MCP Observability **MCP Observability** gives your AI agents—and you—a single HTTPS endpoint to explore logs, metrics and traces for any application, without learning half-a-dozen vendor APIs. • Standards-based: OpenTelemetry ingestion, Grafana/Loki/Tempo storage • AI-ready: implements the Model-Context-Protocol (MCP) so agents can ask natural-language questions about performance issues • Zero-lock-in: self-host via Docker Compose, Kubernetes (Helm) or any cloud container platform --- ## 1 · Why would I use this? Imagine asking your favourite coding agent: > "The /checkout endpoint went 💥 at 09:33 UTC. Why?" The agent queries the MCP endpoint, pulls traces, error logs and latency metrics, and replies with an RCA—and a pull-request suggestion. **That** is what this repo enables. --- ## 2 · How do I run it? Choose the deployment style that fits you: | Environment | Quick start | Full guide | |-------------|------------|------------| | Local laptop | `docker compose -f mcp-obs.yml up -d` | [Docker guide](docs/guides/docker.md) | | Kubernetes | `helm install mcp charts/mcp-obs` | [Kubernetes & Helm](docs/guides/kubernetes.md) | | Cloud (ECS, Cloud Run, Azure CA) | see Terraform/CLI snippets | [Cloud deployment](docs/guides/cloud-deployment.md) | After the stack is up: ```bash export MCP_TOKEN="$(openssl rand -hex 16)" # use the same token you passed during install # Health check curl -k -H "Authorization: Bearer $MCP_TOKEN" https://<HOST>:8000/health ``` Open Grafana at `https://<HOST>:3000` (admin / $GF_ADMIN_PASSWORD) to browse dashboards. --- ## 3 · Talking to the MCP endpoint All requests are HTTPS + Bearer-token. Key endpoints: | Path | What it's for | |------|---------------| | `/logs/errors?limit=100` | Latest error logs from Loki | | `/metrics/latency?percentile=0.95` | 95-th percentile latency from Prometheus | | `/resources` | Metadata describing the data sources your agent can query | | `/prompts` | Parameterised prompt templates you can re-use | ### Example agent prompt > "Retrieve the top three slowest routes over the last hour and suggest an optimisation." that becomes an MCP `SamplingRequest` under the hood; the server stitches data from Prometheus & Tempo and returns JSON your agent can read. --- ## 4 · Instrumenting your application Pick your language guide and drop in the ready-made snippet: * Python & FastAPI – [docs/observability/python_fastapi.md](docs/observability/python_fastapi.md) * Java / Spring Boot – [docs/observability/java_spring.md](docs/observability/java_spring.md) * Node.js – [docs/observability/nodejs.md](docs/observability/nodejs.md) * Go – [docs/observability/go.md](docs/observability/go.md) * Next.js – [docs/observability/nextjs.md](docs/observability/nextjs.md) * React SPA – [docs/observability/react.md](docs/observability/react.md) Need only a subset (e.g. logs-only)? Use the [Partial-deployment guide](docs/guides/partial-deployment.md). --- ## 5 · Security in two minutes 1. **Bearer token** – set `MCP_TOKEN` in your deployment and pass it in `Authorization` headers. 2. **TLS** – self-signed by default (dev), or plug in cert-manager / ACM / Cloud CA. 3. **mTLS (optional)** – enable client cert auth by mounting your CA & toggling `security.mtls=true` in `values.yaml`. --- ## 6 · Troubleshooting cheatsheet | Symptom | Check | Fix | |---------|-------|-----| | `401 Unauthorized` | Correct token? | Pass `-H "Authorization: Bearer $MCP_TOKEN"` | | No traces showing | Otel-collector healthy? Endpoint URL correct in client? | Verify port 4318 reachable | | Grafana empty dashboards | Components disabled | See [Partial deployment](docs/guides/partial-deployment.md) | --- ## 7 · I want to hack on the code! Fantastic 🎉 — head over to **[docs/contributing](docs/contributing/)** for architecture diagrams, dev-environment setup, and the contributor workflow. The README you're reading will stay laser-focused on **users** and **agents**.