Debugging a plan in the Graph workbench
Why graph
- Deterministic by default. Steps are plain tool calls able to reference past step output. 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.
- Any tool you can imagine. Native MCP support and an intuitive tool framework brings every integration you could possibly need to your fingertips.
- 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.
Prototype in your favorite harness, test and verify in the workbench
Graph’s built-in MCP server brings the entire development tool set and execution runtime to your coding harness of choice. Coupled with Graph’s pre-built skills, your agent will have the tools it needs to quickly iterate on every workflow task you throw at it. Once you’re happy, open the plan in the graph workbench and run a full end-to-end test to make sure everything is running correctly.Where to go next
Quick start
Install, scaffold a project, run your first plan.
Core concepts
Plans, the tool catalog, the shape cache, the workbench, and what a run costs.