MCP 服务器

Observability Agent MCP

A portable, read-only Model Context Protocol server for turning observability data into bounded evidence that AI agents can inspect safely.

README

Pak Satpam

The bounded observability and CI guard for AI agents

Pak Satpam gives an AI agent the evidence it needs to understand infrastructure health, investigate incidents, inspect CI failures, and show Grafana visuals without handing the model a shell or unrestricted infrastructure access.

</div>

Pak Satpam means the security guard. It watches, reports, and follows a strict access policy. It does not become the infrastructure administrator.

This is the production-ready evolution of the original Pak Satpam prototype: a portable Model Context Protocol server with deterministic provider adapters, bounded responses, redaction, and approval-gated CI operations. The agent and chat experience remain separate, so the same MCP can serve Hermes/Tabby, desktop agents, CI assistants, or another MCP-compatible client.

What It Does

Engineer
   |
   v
AI agent (Hermes, Tabby, desktop client, or another MCP client)
   |
   |  stdio or authenticated private HTTP
   v
Pak Satpam
   |-- validates every request against a strict schema
   |-- bounds queries, time windows, output size, and labels
   |-- normalizes and redacts provider evidence
   |
   +--> Grafana panels and dashboard PNGs
   +--> Prometheus / VictoriaMetrics metrics
   +--> VictoriaMetrics alert state
   +--> GitHub Actions evidence

The default server exposes seven read-only observability tools. An optional CI module adds four read-only tools and one tightly scoped operation that can rerun failed GitHub Actions jobs only after a fresh, one-time operator approval.

Pak Satpam does not run an LLM, receive chat messages, execute arbitrary shell commands, modify source, deploy workloads, read secrets, or silently expand its own permissions.

Tool Surface

Observability

Tool	What the agent receives
`observability.capabilities`	Configured providers, features, and safety limits
`observability.health_snapshot`	Bounded service and scrape-target health
`observability.active_alerts`	Normalized active-alert metadata
`observability.query_metrics`	Allowlisted instant or range metrics results
`observability.render_panel`	One allowlisted Grafana panel as PNG evidence
`observability.render_dashboard`	One allowlisted Grafana dashboard as PNG evidence
`observability.incident_context`	A compact evidence bundle for an alert or service

CI/CD (optional)

Tool	What the agent receives or may request
`ci.workflow_status`	Status for one allowlisted workflow run
`ci.failed_job_analysis`	Deterministic failure classification
`ci.log_evidence`	Bounded and redacted job-log evidence
`ci.remediation_plan`	A runbook-backed dry-run plan
`ci.rerun_failed_workflow`	Approved rerun of failed jobs only

The CI module stays disabled until the deployment provides repository and workflow allowlists, a GitHub App identity, a replay-safe approval key, and metadata-only audit storage. The public .github/workflows/goal14-controlled-fixture.yml workflow provides a bounded failure-and-rerun test. See the CI/CD runbook.

Run It

Pak Satpam requires Node.js 22 or newer.

npm ci
npm run build
node dist/cli.js

The command speaks MCP over stdio. Configure it in a compatible client instead of typing into the process directly:

{
  "command": "node",
  "args": ["/absolute/path/to/pak-satpam/dist/cli.js"]
}

Run the complete local verification suite with:

npm run validate

Container

The public non-root image supports linux/amd64 and linux/arm64:

ghcr.io/hamardikan/observability-agent-mcp

Production deployments should pin the image by immutable @sha256: digest. For a local build and stdio smoke run:

npm run container:build
docker run --rm -i observability-agent-mcp:local

Connect It to an Agent

Client location	Recommended transport	Intended use
Same machine	stdio	Desktop and CLI agents
Private network	Streamable HTTP	Shared Hermes/Tabby or agent runtime
OCI host	stdio or private HTTP	Podman/Docker deployments
Public network	Not ready	Requires OAuth and tenant isolation first

Private HTTP mode uses a file-injected bearer credential and an exact Host allowlist. It is designed for a private, single-operator network. Publishing the repository or image does not make an unauthenticated public endpoint safe. See Client compatibility and the Security model before deployment.

Visual Evidence

Grafana visuals are first-class MCP evidence. Panel and dashboard tools return PNG ImageContent together with structured metadata: source, observation window, dimensions, byte size, freshness, truncation, and warnings.

Rendering is opt-in. Normal health and metrics requests do not spend browser, renderer, or image-context resources. If rendering is unavailable, Pak Satpam fails to a structured evidence response instead of inventing a graph.

Why Not Just Use Grafana MCP?

The official Grafana MCP is the right choice for broad Grafana-native queries and administration. Pak Satpam owns a narrower boundary intended for operational agents:

Capability	Grafana MCP	Pak Satpam
Grafana administration	Primary owner	Not implemented
Grafana datasource queries	Broad support	Narrow allowlisted adapter
Direct Prometheus-compatible backend	Secondary path	Supported
Provider-neutral incident evidence	Provider-specific	Primary contract
Conservative read-only default	Configurable	Required
Approval-gated CI evidence	Not its scope	Optional module

Both servers can be offered to one agent, but every request must have one clear owner. Pak Satpam never silently delegates to another MCP server.

Project Boundary

This public repository owns the portable protocol, schemas, provider adapters, redaction, tests, npm package, and OCI image. A deployment repository should own private endpoints, network policy, provider allowlists, credentials, and runtime configuration. Private topology and secrets do not belong here.