Coding agents don't lack intelligence — they lack bounded architectural context. arcade-agent is the map: components, dependencies, smells — recovered from your code, served over MCP, versioned in your repo.
This run is the tool analyzing its own repository at v0.3.0 — the same baseline its CI re-commits on every merge.
Tree-sitter parsers extract entities and references. Java and Python out of the box; C, TypeScript, Go, Kotlin and Rust with the languages extra.
Clustering algorithms rebuild components, dependencies, cycles and coupling.
An MCP server exposes the model as budget-aware tools any coding agent can call mid-task.
claude-code I'll check the blast radius before refactoring. → diff_impact(changed_files=["src/auth/session.py"])
arcade-agent { "affected_components": ["auth", "api-gateway"], "num_downstream": 14, "broken_contracts": [] }
claude-code Two components affected, no broken contracts — proceeding with the refactor.
o200k_base tokenizer; Claude's tokenizer counts 1.09–1.25× as many. The full graph costs about as much as the source — the savings come from summaries and budgets. Table and caveats.$ arcade-arch-diff --source . - component "billing" depends on: [db] + component "billing" depends on: [db, auth] new cross-component edge → posted as a PR comment
arcade-arch-diff --update-baseline writes .arcade/baseline.json on every merge; the same command diffs against it on every PR. No SaaS, no account — a file and a markdown comment, informational by default, same as arcade-self-analysis.Why a diff on every PR matters: Decay Front, a four-minute film on how architecture drifts one commit at a time.
Coming in v0.4.0: a guardrail that blocks. Write the intended architecture down in an architecture.spec.json and arcade-guard check fails the commit or the CI job on a violation; agents get the same check as MCP tools while they write. On one greenfield task with Claude Haiku 4.5 (20 runs per condition), the spec alone in the repository left 5/20 runs violating it; with the guardrail tools, 0/20 (one-sided Fisher p = 0.024). Most of the effect came from the in-loop check catching violations the agent then repaired (12/20 runs). One task — evidence the mechanism works, not how often it matters.
Five clustering algorithms — pkg, wca, acdc, arc, limbo — tunable, deterministic output.
Cycles, hubs, coupling and instability metrics per component.
context_for_task, diff_impact, api_surface, dependency_cone, the five guardrail tools (v0.4.0) — plus 15 more, all budget-aware.
Every tool respects a token budget — built for agent context windows.
Chain tools without re-serializing full graphs between calls.
One GitHub Action, one committed baseline, a markdown diff on every PR.
Prefer finished reports over raw tool calls? The arcade-analyze skill runs the whole pipeline from a natural-language request — "analyze the architecture of this repo" — and opens an interactive report: pan/zoom component diagram with drill-down, dependencies + DSM, ranked recommendations, and an animated flow simulation.
Live demo, no install — the animation links to the real interactive visualizer, generated from arcade-agent's own codebase. Setup: docs.