The open infrastructure stack for governed AI systems.
Every repository does exactly one job, follows one philosophy, and composes into one verifiable architecture — organized along the agent lifecycle.
Governance, compliance, and audit are all downstream of one primitive: a
canonical, hashable, tamper-evident unit of what happened that any party can
verify without trusting the system that produced it. That primitive is
octopus-evidence — and it is
not just a shared library, it is a real dependency edge. Across the stack, the
producers of facts project them into the same Evidence atom:
octopus-evidence
(canonical hash · tamper-evident chain · keyed integrity)
▲ ▲ ▲ ▲
Replay │ Observe │ Inspect │ Runtime │
(transcript (observation (each finding (each autonomy
hashes) → evidence) → evidence) decision → evidence)
Hand an auditor a JSON export and one command — octopus-evidence verify — and
they independently confirm what an agent did, what governance saw, and why it
was allowed to act, without running or trusting any of our stores. That is the
EU AI Act Art. 12/14 logging story, in shipping code.
| Stage | Repo | One job | On the spine |
|---|---|---|---|
| Primitive | Evidence | The canonical, verifiable unit of what happened. | root |
| Collect | Scout | Collect evidence, not webpages. | |
| Observe | Observe | Turn untrusted events into trusted observations. | → evidence |
| Understand | Experience | Knowledge is earned, not stored. | |
| Coordinate | Blackboard | Shared cognition for coding agents. | |
| Track | Workstate | Where work is, and why it moved state. | → evidence |
| Act | Runtime | Make unsafe actions structurally impossible. | → evidence |
| Verify | Replay | Reproduce every agent incident byte-for-byte. | → evidence |
| Govern (every stage) | Inspect | Governance lint for AI workspaces. | → evidence |
Collect → Observe → Understand → Coordinate → Act → Verify
Scout Observe Experience Blackboard Runtime Replay
Inspect governs every stage
octopus-evidence is the atom they all ride
New to the stack? Blackboard is the entry point — shared, local-first memory for your coding agents, zero backend, one command to wire into Claude Code / Cursor / Codex:
npx octopus-blackboard quickstart # init + detect client + paste-ready MCP configLive on npm today — Apache-2.0, zero third-party dependencies, published with provenance:
npm i octopus-evidence # the Evidence primitive → auditor CLI: octopus-evidence verify export.json
npm i -g octopus-inspect # governance linter → octopus-inspect . --format sarif|evidence
npm i octopus-runtime # governed execution — every autonomy decision → verifiable evidence
npm i octopus-observe # untrusted events → trusted observations → verifiable evidence
npm i -g octopus-scout # governed web/PDF ingestion — semantic search works offline, no API key
npm i -g octopus-blackboard # shared agent memory (see quickstart above)
npm i octopus-workstate # verifiable work-state ledger — a task is a provenance chain, not a to-do
npm i -g octopus-replay # determinism harness → octopus-replay --helpDrop Inspect into any repo's CI in two steps — findings land in the Security tab:
- uses: octoryn/octopus-inspect@v0.3.2
with: { path: ".", fail-on-findings: "false" }
- uses: github/codeql-action/upload-sarif@v3
with: { sarif_file: "${{ steps.inspect.outputs.sarif-file }}" }- Evidence is the root — everything reduces to a verifiable record — Evidence
- Observation precedes reasoning — Observe
- Trust must be earned — Experience
- Unsafe execution must be impossible — Runtime
- Every action must be reproducible — Replay
- Governance begins before runtime — Inspect
- Agents think together — Blackboard
- octopus-linkedin — a governance example: draft → review → approve → publish, the stack's discipline applied to a real outward action.
- octopus-agentos — an enterprise AI development environment (the stack's distribution form).
Open · Composable · Apache-2.0 · zero third-party dependencies. No black boxes, no SaaS lock-in — built to be forked, audited, and composed into your own systems.