arcade-agent recovers an architecture model from source — components, dependencies, smells, metrics — and serves it over MCP as 19 tools. These are the three jobs that model is actually for at v0.3.0. Each page names the exact tools involved and what they return.
Your agent is asked to change code it has never read. Instead of paging files into the context window until something fits, it asks three task-shaped questions and gets a ranked, capped answer back.
context_for_taskapi_surfacedependency_coneBefore the edit: which entities, components and downstream dependents does this file touch, and does it break a public contract. After the merge: a committed baseline that gets diffed on every pull request when the recovered architecture changes shape.
diff_impactarcade-arch-diffcompareNobody remembers what the components are, and the diagram in the wiki is three years old. Recover the architecture from the code itself, see which components carry cycles or overload, and get numbers you can re-measure after the refactor.
recoverdetect_smellscompute_metricsSame model, same install, no service in the middle:
pip install "arcade-agent[languages,mcp]" — Python 3.12+. Java and Python parse out of the box; C, TypeScript, Go, Kotlin and Rust come with the languages extra, and arcade-mcp needs the mcp extra to start.max_tokens and degrades its response instead of erroring — see token budgets.ingest → parse → recover hand back session IDs that later calls reference, so a graph is built once per session (or in one call, with analyze).Install, point your MCP client at arcade-mcp, and run the first analysis — the getting-started page walks a real captured transcript against pallets/click.