Skip to main content
graph is a command-line workflow engine for AI-augmented operations, CI, automation, and ETL. You describe a workflow as a plan — a small YAML file that chains tool calls with typed dataflow between steps and LLM inference exactly where you place it. Manage your entire workflow code base using the graph workbench. Commit everything to version control. Build custom mcps to scale AI Ops across your teams.
The workbench after a completed weekly_status run: the chat agent on the left, the fully-executed step tree with exit-gate, map, and solver rows on the right, and the map's dataflow template in the detail pane.

A report plan run end to end in the workbench: an exit gate for the quiet weeks, a per-file map, and the solver's report — the same steps in the same order, every Monday.

An agent gives you a different answer every run. A plan is an artifact — versioned in a repo, reviewed in a PR, and cheap enough to schedule. That is the difference between a demo and a workflow you can run a thousand times.

​
Why plans

  • Deterministic by default. Steps are plain tool calls with no inference between them. The LLM enters only where you put it: a prompt tool as a step (classify, extract, summarize mid-plan) and a finish that is a written report, structured JSON, or nothing at all.
  • Repeatable and reviewable. A plan is a file you check into a repo, review in a PR, and share with your team. Your Monday status report is the same report every Monday.
  • Composable. Every plan joins the tool catalog as plan__<name>, callable like any other tool — from the CLI, from a script, or from inside another plan. And a plan’s own steps draw from that same catalog: MCP tools, wrapped CLIs, saved queries, prompt tools, and other plans.
  • Bring your own models and storage. Any MCP server; Anthropic, OpenAI, or OpenAI-compatible/local models, assignable per role. A single binary with plain-file storage — or in-memory storage for ephemeral CI runs. Your data stays on your machine.

​
Built for scaled AI Ops

The same plan file works in a terminal, a cron job, and a CI pipeline: finish modes for reports, structured JSON, or silent side effects; meaningful exit codes; ${ENV} secrets that fail loudly; repo-carried config under ./.graph/. See the cookbook.

​
Prototype in chat, review in the workbench

graph ask and graph chat run an agent loop over the same tool catalog — the conversational surface for probing a tool’s real output and prototyping a workflow before you freeze it into a plan (Chat & ask). And plans themselves have a dedicated surface: the workbench (graph wb plan), a dual-pane TUI where the agent drafts a plan into a side pane and you inspect, validate, and step through it with a debugger (Workbench).

​
Where to go next

Quickstart

Install, scaffold a project, run your first plan.

Core concepts

Plans, the tool catalog, the shape cache, the workbench, and what a run costs.

Author a plan

From blank file to running plan, using a real example.

Cookbook

CI checks, reports, and automation — worked examples, annotated.

Related topics

Changelog