Cairn

An agentic incident-analysis copilot. Ask "why did checkout latency spike at 3am?" in plain English; Cairn queries the observability stack, correlates against the deploy timeline, retrieves the relevant runbook, and proposes a root cause with evidence — and can execute a remediation, behind a human approval gate.

This repository implements docs/architecture.md.

The dashboard

The companion npm package (npx @nouman-amjad/cairn dashboard) serves a live local dashboard over any project — no build step, no dependencies, no network.

<details> <summary>Findings, logs, services and chat</summary>

</details>

The four commitments, and where they live in the code

Commitment	Where it is enforced
Tools are the product. Every capability is an MCP tool; the agent has no privileged path to any backend.	services/cairn-mcp-*, packages/cairn-mcp-kit
Inference is a routed resource. A cost-and-sensitivity router splits work between a local 8B and a frontier model.	routing.py
Write actions are never autonomous. A durable approval state machine with idempotency keys and an append-only audit log.	service.py
The eval harness is a first-class service. 30 scenarios with ground-truth causes and 7 gated metrics.	services/cairn-eval

Layout

packages/
  cairn-core/          domain model, config, auth, DB, sensitivity, artifacts
  cairn-mcp-kit/       MCP scaffolding: identity, OPA guard, result capping, versioning
services/
  cairn-gateway/       OIDC, rate limits, cost budgets, circuit breaker, SSE fan-out
  cairn-orchestrator/  the agent loop as a state machine persisted to Postgres
  cairn-router/        model routing, cost accounting, vLLM + Anthropic clients
  cairn-approval/      approval state machine and the Slack gate
  cairn-mcp-observability/  metrics, logs, traces, deploys, artifacts
  cairn-mcp-runbooks/       hybrid search, ingest, past-incident recall
  cairn-mcp-actions/        approval-gated write tools
  cairn-eval/          30 scenarios, 7 metrics, the CI gate, a seeded stack
  cairn-cli/           `cairn ask "why did checkout spike?"`
ui/                    Next.js: chat, trajectory viewer, approvals
cairn-deploy/          Helm chart, ArgoCD app-of-apps, OPA bundle, prompts
cairn-infra/           Terraform: VPC, EKS, RDS, Karpenter, IRSA, S3
docker/                one Dockerfile for all Python services, plus vLLM
docs/adr/              14 architecture decision records

cairn-deploy and cairn-infra are vendored here for review. In production they are separate repositories with separate lifecycles — see ADR-013.

Getting started

make install     # uv sync + npm ci
make up          # Postgres (pgvector), Redis, MinIO, OPA
make migrate
make test        # 246 tests
make selfcheck   # 21 module self-checks
make eval        # 30 scenarios through the real agent loop

No GPU and no API key are needed for any of the above. make eval runs in heuristic mode, which exercises the whole pipeline — loop, tool capping, persistence, the seven metrics, the gate — with a scripted stand-in instead of a model. See the caveats.

To run against real models, set CAIRN_ROUTER_ANTHROPIC_API_KEY and point CAIRN_ROUTER_VLLM_URL at a vLLM server.

Debugging with an MCP client

Every MCP server speaks stdio as well as Streamable HTTP:

make mcp-stdio
# or, the way an outside user would:
uvx cairn-mcp-observability --stdio

Point Cursor, Zed or any other MCP client at it and you see exactly what the agent sees. That costs about twenty lines (ADR-002) and pays for itself the first time a tool result looks nothing like you expected.

What is and is not verified

Being specific, because "it works" is not a claim worth making vaguely.

Verified by running it here:

• 246 tests pass against a real PostgreSQL 16 + pgvector, including the approval-gate safety properties (no self-approval, no double execution, no execution without approval, an append-only audit log enforced by a database trigger) and the router property that restricted data never reaches a cloud model — checked exhaustively across every task class and tier state.
• 21 module self-checks pass.
• The Alembic migration applies cleanly to a real database.
• 12/12 OPA policy tests pass.
• The 30-scenario eval runs end to end through the real agent loop; all seven metrics are above target and the gate correctly blocks both a simulated regression and a below-target run.
• The UI typechecks and builds. The CLI's five commands resolve.

Not verified here, and why:

• Eval accuracy numbers. The committed baseline is heuristic mode, which scores the harness rather than the agent. Real numbers need make eval-record against a live router. The LLM cause judge is built but not calibrated — nobody has labelled 100 runs, so κ is unknown.
• Terraform. No terraform binary in this environment; fmt, validate and tflint run in CI. Nothing has been applied to an AWS account.
• Helm rendering. helm lint, template and kubeconform run in CI for all three environments. They were not run locally.
• Every cost figure in docs/cost.md. The accounting that would measure them is built and tested; the traffic to populate it does not exist.
• The vLLM performance table. Arithmetic from memory bandwidth, not measurement. docs/inference.md gives the benchmark command and a falsification threshold.
• Chaos testing. Phase 7 of the roadmap has not been run.

docs/roadmap.md tracks this phase by phase.

Documentation

• docs/architecture.md — the design this implements
• docs/adr/ — 14 decision records, each with its consequence
• docs/security.md — threat model and where each control lives
• docs/inference.md — GPU sizing, KV-cache arithmetic, vLLM flags
• docs/cost.md — the cost model, including whether the GPU pays for itself
• docs/operations.md — runbook: what pages, what to do
• docs/roadmap.md — phases, risks, and what is actually done
• services/cairn-eval/README.md — how to read a gate failure