Occasio

Occasio

Local-first audit chain and identity gate for AI coding agents. Your prompts, your tool calls, your audit log; all on your machine, cryptographically verifiable later — and an agent may request an identity (ssh / cloud / root), not silently assume one.

Occasio is a local proxy that sits between your AI coding agent (Claude Code, Cline, or anything that talks to the Anthropic API) and whichever LLM endpoint you configured. Every tool call passes through one human-readable policy file you control. You can see what is leaving your machine in real time, block what should not happen, and end up with a tamper-evident hash-chained log that a third party can verify offline months later if you ever need to prove what your agent did.

Nothing about Occasio sends data to a third party. There is no Occasio cloud, no Occasio account, no telemetry. Read docs/WHY-LOCAL.md for the architecture, docs/COMPARE.md for how this differs from cloud-hosted AI observability tools, and docs/SUSTAINABILITY.md for how a local-first product is funded under Apache 2.0.

[occasio eyes — live browser view of an outbound exchange, with redacted secrets highlighted in the tool result]

Screenshot: occasio eyes --demo against synthetic data — no real paths, no real traffic. The "(demo)" badge top-left is always shown in demo mode.

npm install -g @occasiolabs/occasio
occasio demo audit       # Auditor scenario: signed attestation + cross-verifier proof (10s, no API key)occasio demo attest      # End-to-end attestation pipeline against a synthetic chainoccasio demo anomalies   # Live EDR detection on a synthetic adversarial chainoccasio harness          # Real Claude Code attacking a denied path — defense holds

The first three demos run against synthetic data so you can see the full pipeline in seconds with no external dependencies. The fourth spawns a real Claude Code subordinate under your Anthropic login (bundled auth — no API key required) and proves the defense end-to-end. Start with demo audit — it answers the only question that actually matters: "prove what your AI agent did in CI."

Identity gate — request, don't assume

The incident this was built for: an agent is asked only for a deploy command, but on its own it sshes into the server and reads env. The strict policy stops exactly that — an AI agent may request an identity, it may not silently assume one.

• Exfiltration is denied, tool-agnostically. printenv, cat .env, /proc/self/environ, private-key reads, grep-for-secret-names — blocked no matter which command or tool reads them; the output never reaches the model.
• Identity borrows need a human. ssh / scp, az / the cloud control plane, sudo / systemctl — a fail-closed BLOCK with a "requires human approval" refusal the agent cannot satisfy on its own.
• The handshake distinguishes who is who. The agent requests; you authorize once, out-of-band, from your own terminal:

occasio init --template strict          # the identity-gated posture# agent runs `ssh deploy@host`  →  BLOCKED, pending apr_…occasio approvals approve apr_… --once  # you, single-use, short TTL# the agent's retry passes through once — then it's blocked again

The agent cannot self-approve: the approval control plane is in its deny-zone and the token is HMAC-signed. The chain records actor=ai_agent · delegator=you · approved_by=you. A second, non-proxied enforcement point — a PreToolUse hook (occasio hook --install) — covers execution that doesn't pass through the proxy.

Full design, threat model, and the honest residuals (runtime indirection, egress): docs/identity-gate.md.

Two ways to use it

Most people land here for one of these — pick the column that fits and skim accordingly.

Both views read the same underlying log. You don't have to pick one — running the proxy gives you both for free.

Quickstart

New here? The 5-minute Getting Started guide walks the identity-gate flow end to end — install → init --template strict → run your agent → approve a borrow → verify the log. Every command in it is tested against the published release.

Requires Node.js ≥ 18. Works on Windows, macOS, Linux.

npm install -g @occasiolabs/occasio   # Installoccasio doctor                          # Verify setup (Node, claude CLI, port, profile)occasio init                            # Write ~/.occasio/policy.yml (dev-default)occasio init --template strict          # …or the identity gate: deny secrets, gate ssh/cloud/root behind approvaloccasio register                        # Add 'claude' shell alias (one-time)claude "read package.json and tell me the version"

After the alias is registered, every claude invocation routes through Occasio transparently. Audit-chain rows accumulate at ~/.occasio/pipeline-events.jsonl.

Inspect the run:

occasio status                  # Session totalsoccasio replay --detail         # Run-level auditoccasio audit verify            # Re-walk the hash chain end-to-endoccasio anomalies               # Run EDR detectors over the last 15 minutesoccasio attest --run-id <uuid>  # Build a behavioral attestation for one sessionoccasio explain <event-id>      # Why was this blocked? — the matched policy rule + how to unblock

Command matrix

Grouped by what you're doing. (stable) = load-bearing with test coverage; (beta) = works end-to-end, narrower; (alpha) = scaffold. Full list: occasio help.

Evidence workflow

Produce one file an auditor or CI receives and checks in a single command:

occasio attest --run-id <uuid> --sign          # behavioral attestation (Sigstore optional)occasio bundle --run <uuid> --out run.occasio.json   # pack attestation + chain slice + policy + manifestoccasio verify run.occasio.json                 # 6 offline checks; exit ≠ 0 on tamper

verify checks the schema, the manifest hashes, the chain slice integrity, the policy binding, the git-state vs chain cross-check, and the Sigstore signature when present — all against data embedded in the file (never the producer's machine). The run is bound to the concrete code it touched via git_state rows (HEAD + diff hash + changed files). Signing is optional; an unsigned bundle still verifies everything else. Bundles embed absolute producer paths — an internal-audit artifact; review before sharing publicly. → docs/VERIFY.md

Policy workflow

~/.occasio/policy.yml governs the Occasio-controlled path: secret handling, deny_paths/allow_paths, custom deny_patterns, per-round limits, and tool routing. Approve it and pin against drift:

occasio policy init                              # starter policy.ymloccasio policy validate                          # lint before you rely on itoccasio policy lock --sign --out policy.lock.json   # record the approved policyoccasio policy diff --since policy.lock.json     # CI gate: exit 1 if the policy drifted

→ docs/POLICY.md

Scanner & preflight

occasio scan --file .env                         # explainable secret findings (masked, exit 1)occasio preflight simulate --read ~/.ssh/id_rsa --bash "npm test" --strict   # would the policy block this?

scan never prints the secret in plaintext (masked snippet + SHA-256). preflight simulate runs candidate actions through the same policy engine the runtime uses and shows the matched rule + how to unblock for anything it would block. → docs/SCAN.md · docs/PREFLIGHT.md

Verify the local-first claim

occasio doctor --paranoid

Scans the installed source for every outbound network primitive, classifies each callsite, checks for telemetry SDK signatures, and surfaces the audit chain status. Exits non-zero if any critical finding appears. JSON output via --paranoid --json. See docs/WHY-LOCAL.md for the architecture this verifies.

Where Occasio sits alongside SLSA Provenance and CycloneDX AI-BOM: docs/SUPPLY-CHAIN-TRIANGLE.md. Occasio is the runtime behavioral leg.

Live visibility — for developers

The audit chain and signed attestations answer "what did the agent do, prove it" — the auditor's question. Sometimes you also want the simpler one: "what is the agent doing right now?" What is leaving your machine in this HTTP request. Which files have already gone to Anthropic in this session. What the system prompt looks like. What got redacted before it shipped.

occasio eyes is a local browser UI on http://127.0.0.1:3002 that shows exactly that. Capture is opt-in via --eyes on the proxy; nothing leaves the machine, all storage stays under ~/.occasio/eyes/.

occasio eyes --demo            # synthetic data, no proxy needed (10-second tour)occasio claude --eyes          # then in another terminal:occasio eyes                   # browser tab opens automatically

What you see:

• Sidebar grouped by user prompt — one "read the readme" expands into the 3–6 HTTP round-trips it actually took (Claude Code's agentic loop made visible).
• Session sparkline + per-file aggregate — outbound size over time, which files have been sent to the cloud and how often.
• Per-exchange byte-decomposition — stacked bar showing system / history (replay) / new-this-turn / tools-framing. The "small" request that's 124 KB is suddenly explainable: 108 KB of it is the same system prompt every time.
• Tabs: Request · Response · Tools (full local tool outputs — Read/Glob/Grep/Bash bytes that the interceptor handled locally and never sent up) · Diff (pre-transform vs sent — see the literal secrets that were redacted, in the clear, locally) · Headers · Raw SSE bytes · Stats.
• Live flash overlay when a new outbound lands — useful both for understanding and for demos.

Screencast-safe view — --sanitize

Recording a demo or screenshot of occasio eyes against a real session normally leaks your identity: home path (C:\Users\<you>\...), OS username, git email, real name, hostname. The --sanitize flag replaces those with deterministic pseudonyms in the display only — disk contents under ~/.occasio/eyes/ are unchanged so your audit trail stays real.

occasio claude --eyes --sanitize    # capture as normaloccasio eyes --sanitize             # view with identity scrubbed

In the browser UI a cyan dot and (sanitized) badge confirm it's active. Paths like C:\Users\<you>\Desktop\proj become /home/user-7c/Desktop/proj, stable within a session.

What --sanitize covers: $HOME paths, OS username, git user.email / user.name, hostname, and identity-carrying env vars (USER, USERNAME, LOGNAME, HOME, USERPROFILE).

What it does not cover — review before sharing:

• Project paths outside $HOME (e.g. D:\Work\Acme\…)
• Git remote URLs in tool outputs (github.com:org/repo leaks the org name)
• File contents — your name in a comment, commit message, or README is not auto-detected. Sanitize keys off identity values it can discover from the OS and git, not arbitrary substrings.
• The Claude Code TUI banner ("Welcome back X", organization line) — that prints before any HTTP traffic and bypasses Eyes entirely. Crop it from the screencast manually.

The flag is a display filter, not a recording mode. The same Eyes capture can be replayed unsanitized later (just run occasio eyes without the flag).

Dashboard vs. Eyes

Both run as local browser UIs but answer different questions:

Both live on 127.0.0.1 only. No CORS, no auth, no external network.

Commands

Session-level overrides on top of policy.yml:

What it does, in four layers

Under the hood, four layers do the work — Layers 1–2 every run, Layers 3–4 when you ask for an attestation or run the detectors.

Layer 1 — Tool-call interception. A local proxy sits between the agent and the Anthropic API. Read, Glob, Grep, TodoRead/TodoWrite run in-process on your machine; the file bytes never enter the outbound request. A curated set of read-only shell commands (git status, git log --oneline -N, with or without git -C <path>, plus echo / cd cwd-prefix chains) are also executed in-process. Other shell reads (cat <file> and similar) are policy-analyzed for embedded read paths so deny_paths enforces consistently, but the command itself executes server-side via the Bash tool.

Layer 2 — Policy enforcement. Every tool call hits one decision: LOCAL / PASS / BLOCK / TRANSFORM, driven by policy.yml. deny_paths is enforced on the realpath-resolved absolute path so symlinks and traversal variants resolve to the same denial. block_secrets_in_tool_results redacts API keys and JWTs out of any tool output before it re-enters the prompt. Hot-reload: edits to policy.yml take effect on the next call, with a policy_loaded row written to the audit chain. The strict template adds the identity gate — deny_commands block exfiltration behaviours (env dumps, secret-name greps) and identity_approval gates identity borrows (ssh / az / sudo) behind a single-use, human-approved token (see Identity gate).

Layer 3 — Behavioral attestation. occasio attest --run-id <uuid> produces a self-contained JSON predicate that commits to the full audit-chain slice for one agent session: every tool call, every block, every transform, every redacted secret, plus the active policy's SHA-256 hash and rules digest. --sign wraps it in an in-toto Statement v1 and Sigstore-signs it using GitHub Actions OIDC (no key management). The predicate type URI is agent-attestation/v1. Two independent reference verifiers ship — Node (occasio attest verify) and Python (docs/attest_verify.py) — and the test suite asserts they agree byte-for-byte on the same payload.

Layer 4 — Anomaly detection (EDR). occasio anomalies runs four detectors over a time window of the audit chain: deny-rate spike, file-read-volume burst, previously-unseen tool-input shape, secret-redaction-rate spike. Severity escalates against your historical baseline — roughly a ×10–×20 ratio above normal triggers HIGH at the detector level; against a sparse normal baseline the observed multipliers in practice land between ×100 and ×1000. See docs/edr-demo.md for the reproducible defense-in-depth walkthrough.

Architecture

agent (Claude Code / Cline / MCP / Computer Use)  │  ▼  tool call┌──────────────────────────────────────────────────────────────┐│  Occasio proxy                                            ││                                                              ││  Layer 1: adapter parse → canonical event                    ││  Layer 2: policy decision (LOCAL / PASS / BLOCK / TRANSFORM) ││  Layer 2: deny_paths + deny_patterns + secret redaction      ││  Layer 2: native dispatch for LOCAL/TRANSFORM tools          ││           ──► row appended, SHA-256-chained                  ││  Layer 4: anomaly detectors (windowed, on-demand or live)    │└──────────────────────────────────────────────────────────────┘  │  ▼  cloud-bound: only PASS calls, with shaped result if TRANSFORMAnthropic API
End of session  │  ▼  occasio attest --run-id … --signLayer 3: signed in-toto Statement → Sigstore bundle → GitHub Check Run  │  ▼  independent verifierNode / Python / cosign — all must agree

Verification

Three independent checks, all required for a verified attestation:

1. Sigstore signature — Fulcio certificate chain + Rekor inclusion proof. Verifiable by any sigstore-conformant tool (cosign verify-blob, sigstore-js, sigstore-python).
2. DSSE payload ↔ attestation predicate equivalence — re-decode the in-toto Statement inside the bundle, canonicalise the predicate via RFC 8785 subset, compare byte-for-byte with the attestation predicate (minus the signature metadata field).
3. Audit chain integrity — SHA-256-walk every prev_hash → hash link from the GENESIS sentinel, then assert the attestation's first_hash and last_hash appear in the chain in the right relative order.

Two reference verifiers ship side by side:

• Node: occasio attest verify <file>
• Python: python docs/attest_verify.py <file> — stdlib + optional sigstore-python, reuses docs/audit_walker.py for the chain step. See docs/python-verifier.md.

Cross-language invariant (asserted in the test suite as xlang: and xlang-float: cases): both verifiers agree byte-for-byte on the predicate-equivalence and audit-chain steps for the same payload, including tamper-detection cases. Non-integer numbers are rejected by both canonicalize implementations so a future schema cannot silently introduce divergence.

The Sigstore signature step uses the standard DSSE-wrapped in-toto Statement format; any sigstore-conformant tool verifies it (cosign verify-blob, sigstore-js, sigstore-python). The test suite mocks the signing path; a real-OIDC end-to-end signed-and-verified round-trip requires a GitHub Actions environment and is exercised by the integrations/attest-action/ workflow in CI.

A third partial verifier runs in-browser at integrations/attest-view/ for drag-and-drop inspection. The browser performs the predicate-equivalence and audit-chain steps but defers Sigstore certificate-chain verification to one of the two CLIs (bundling Fulcio/Rekor trust roots in-browser is intentionally not done; the page is explicit about it).

Why now

Three regulatory drivers, all converging on the same requirement: runtime evidence of AI-agent behavior must be cryptographically verifiable.

• EU AI Act Art. 12 mandates comprehensive automated logging of high-risk AI systems. In effect from 2026 for regulated sectors.
• NIST AI RMF (GOVERN, MEASURE, MANAGE families) is becoming required in US Federal procurement and is influencing FedRAMP AI controls.
• SOC 2 Common Criteria are extending to AI-agent controls — auditors at major firms started asking "show me the agent's tool-call log" in 2026 audits.

There is currently no off-the-shelf product producing a signed, third-party-verifiable artifact for what an AI coding agent did inside your CI. Occasio fills that gap with an open schema (Apache-2.0) and ships the reference implementations for it.

Log format

All data is stored locally at ~/.occasio/:

~/.occasio/  pipeline-events.jsonl        # tamper-evident audit chain (SHA-256 linked)  policy.yml                   # active policy  session.json                 # current run_id, totals  logs/YYYY-MM-DD.jsonl        # per-request log  baseline/<cwd-hash>.json     # per-project behavior baseline (opt-in)  eyes/                        # `occasio eyes` capture (opt-in via --eyes)    payload-NNNNNN.json        # per-exchange metadata + extracted blocks    content/<sha>              # content-addressed blob store (file bytes,                               # tool outputs) — dedup by SHA-256

The audit-chain row schema is documented in docs/AUDIT.md. Each row carries prev_hash and hash (SHA-256 hex), with the first row chained from a fixed GENESIS sentinel (64 zeros). occasio audit verify and docs/audit_walker.py are independent implementations of the walker.

Demos

• EDR defense-in-depth — real Claude Code attacking a denied path under your policy, all blocks held, EDR fires HIGH ×100–×1000 over baseline. Reproducible in <2 minutes.
• Reference Pipeline — PR with AI-agent → GitHub Action signs attestation via Sigstore keyless → Check Run on the PR → independent verification offline.
• Cross-protocol governance — the same deny_paths rule producing identical BLOCK rows under Claude Code's HTTP proxy and the MCP server.

Reference

• spec/agent-attestation/v1/README.md — predicate type specification
• schemas/agent-attestation-v1.json — authoritative JSON Schema
• docs/AUDIT.md — audit-chain row schema and canonical-serialisation rules
• docs/compliance-mapping.md — SOC 2 Common-Criteria mapping
• docs/python-verifier.md — independent Python verifier
• docs/edr-demo.md — defense-in-depth walkthrough
• docs/reference-pipeline.md — end-to-end CI pipeline
• integrations/attest-action/ — GitHub Action that signs + posts a Check Run
• integrations/attest-view/ — static browser viewer for attestation files

Requirements

• Node.js ≥ 18
• Claude Code (npm install -g @anthropic-ai/claude-code) — or any agent that respects ANTHROPIC_BASE_URL
• Python 3 (optional) — required for the independent verifier and LAO context trimming
• sigstore-python (optional) — adds the cryptographic Sigstore step to the Python verifier

License

Occasio is open source under the Apache License 2.0, including an explicit patent grant for safe enterprise use. Versions 0.6.6 and earlier were released under the MIT License and remain MIT in perpetuity for those releases.

Contributions are accepted under Apache-2.0; please sign off your commits per the DCO (git commit -s).