Tool Dependency Graph
An action planner that maps 893 GitHub tools into a 7,376-edge graph
A generator that reads any API toolkit's catalog and infers which actions must run first to supply another's inputs — with schema-backed matching, an LLM pass that can only prune, and 316 automated tests.
- 893 tools mapped
- 7,376 edges inferred
- 84% top-1 accuracy
01
Problem
Agents that call APIs need the order of operations: you cannot comment on a pull request until something has produced its number. That knowledge is buried in schemas and nobody wants to hand-write it for hundreds of tools.
02
What I built
- A generator that reads a toolkit's catalog and infers, per action, which other actions can supply its inputs.
- An interactive viewer for the resulting graph — 893 GitHub tools, 7,376 edges.
03
Technical decisions
- Schema-backed matching does the work; the LLM pass is only allowed to prune candidate edges, never to invent them.
- An independently labelled 40-tool gold set keeps the accuracy number honest.
04
Results
- 84% top-1 accuracy against the gold set.
- 316 automated tests, which caught 11 real bugs during development.
05
What I'd do next
Generalise the catalog reader to OpenAPI specs and MCP tool lists.
Built with
- TypeScript
- Node.js
- OpenAI API