Get running
That's it. Your workflow doesn't change. Your agent calls Decapod before:
- Acting — intent
- Calling the model — context
- Committing — proof
- Touching protected code — boundaries
Decapod is designed to stay out of the human workflow. The agent checks in. You keep talking to your agent like normal.
AI agents do not fail because they lack tools. They fail because they lose intent, skip dependencies, mutate context unsafely, and return vibes instead of proof.
The loop
User
│
▼
Agent ───────┐
│ │
│ ┌────▼────┐
│ │ Decapod │
│ │ (check) │
│ └────┬────┘
│ │
├─────────┤
│ │
Model Agent
│ │
└────┬────┘
▼
User
What Decapod does
- Clarifies intent — What's the goal?
- Bounds context — Only what's needed. Not the whole repo.
- Enforces proof — VERIFIED means gates passed.
- Protects boundaries — No direct writes to master.
Decapod resolves only what's relevant to the user's intent — no context poisoning. Your agent gets surgical context, not the entire codebase.
The constitution
Decapod ships with an embedded engineering constitution.
94 documents covering architecture, security, performance, testing, knowledge graphs, claims, proof surfaces, interfaces, evaluation criteria, and workflows. Everything an engineering org usually keeps in scattered docs, tribal memory, and review culture becomes executable guidance your agent can consult.
Recent research has confirmed what Decapod was built around from the start: AI coding agents waste significant context on irrelevant files. — arXiv:2602.11988
Your agent doesn't guess. It reads the constitution. It cites claim IDs. It follows gates. It produces proof. You just talk to your agent.
Proof lives in the repo
Decapod does not ask you to trust an agent transcript.
Every run leaves its operational evidence in .decapod/:
- captured intent
- resolved context
- generated specs
- todo state
- dependency structure
- boundary decisions
- verification results
- proof artifacts
That directory is the proof surface. It can be inspected locally, reviewed in pull requests, archived with the codebase, and used by the next agent invocation to re-establish state.
No dashboard. No daemon. No hidden memory.
The repo remembers.
What you get
- No daemon.
- No SaaS control plane.
- No hidden agent memory.
- Full operational state stored locally in
.decapod/. - Proof your team can inspect, diff, review, and commit.
Before / After
Before
User: "build auth"
Agent: [full repo in prompt]
→ generates
→ commits
After
User: "build auth"
Agent: [Decapod]
→ intent: auth system
→ context: src/auth/
→ generates
→ [Decapod]
→ proof: verified
→ commits
Running
Use whatever agent you already use: Claude, Codex, Gemini, Cursor.
Guarantees
- Daemonless — runs on-demand
- Repo-native — state in
.decapod/ - Proof-gated — VERIFIED means gates pass
- Boundaries enforced — protected branches locked
Contributing
&&