Skip to content

glassflow/navflow-cookbooks

Repository files navigation

NavFlow Cookbooks

Hands-on, outcome-driven tutorials for NavFlow — the open-source data plane for AI agents. Each cookbook is a self-contained example you run against a real, local NavFlow deployment: stand up a small system, point NavFlow at it, connect an agent over MCP, and measure the outcome.

These are companions to the running product, not a simulation. You bring up navflow up and navflow mcp locally (below), and each cookbook drives them end to end.

New here? → QUICKSTART.md — install NavFlow, run the first cookbook, and read the result in a few minutes.

Prerequisites (shared by every cookbook)

  • NavFlow, installed and running locally — the data plane the cookbooks read through:
    uv tool install navflow      # or: pipx install navflow
    navflow up                   # daemon + console → http://127.0.0.1:8787
    navflow mcp                  # agent endpoint → http://127.0.0.1:8788/mcp   (second shell)
  • Docker (Desktop running) — each cookbook ships its own stack for NavFlow to ingest.
  • uv (or Python 3.11 / 3.12) — to install the cookbook deps.
  • An Anthropic API key — the agent runs are real API calls (default model claude-opus-4-8).

See the NavFlow docs for install, concepts, connectors, and MCP setup.

Cookbooks

# Cookbook Outcome it measures
01 AI SRE — incident response Same incident-response agent, same four production faults, same answers — NavFlow collapses the per-system tool fan-out into one correlated read. Measured in tool calls, turns, and real API cost.

More to come. Each is standalone and ships its own stack.

How a cookbook is laid out

Every cookbook is a directory under cookbooks/ that owns everything it needs — there is no shared framework to learn:

cookbooks/01_sre_incident_response/
  platform/         ← the system this cookbook stands up for NavFlow to ingest (docker compose)
  navflow_client.py ← creates the cookbook's own namespaced NavFlow sources/view/triggers via the API
  *_agent.py        ← the agents under test (a provider-style baseline vs the NavFlow read path)
  run.py            ← run one scenario and print the scoreboard
  report.py         ← run the full benchmark and print the results table
  README.md         ← the walkthrough

Different cookbooks may ship completely different stacks — the shape above is a convention, not a constraint.

Each cookbook is a guest on your NavFlow: it creates its own objects under a namespace prefix (cookbook 01 uses sre_), never touches your existing sources, and cleans up after itself (python run.py teardown).

License

Apache License 2.0 — see LICENSE.

About

Companion cookbooks that take a provider's agent, swap only its read path for one NavFlow query, and measure the difference on a real running system. Includes a self-contained SRE incident-response benchmark — four injectable faults, provider-style tool fan-out vs. one correlated query

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages