The governance flight recorder for AI-written code

Secuarden is a local-first CLI for developers, platform engineers, and security teams who need to investigate AI-assisted changes. It captures evidence from Claude Code and compatible Codex runtimes, so you can inspect agent activity, verify stored records, and connect observations to committed changes.

Install and try it

You need Git, a local Git repository, and an installed, authenticated supported agent. Local capture does not require a Secuarden account.

curl -fsSL https://install.secuarden.ai | sh
cd /path/to/your/repository
secuarden init --repo .

These instructions describe the source on main. Until the 1.0 release is published, the installer downloads the latest published release, which may not include every workflow below. To try this checkout, build from source.

Setup detects supported agents. Choose explicitly with --agent claude-code or --agent codex. Codex requires compatible hooks, configuration, and trust; follow the setup diagnostics. See agent setup.

Now run a real session with your agent in that repository: ask it to read a file, make a small change you can review, and run a relevant command. Then inspect:

secuarden status
secuarden events --last 20
secuarden serve

Open http://127.0.0.1:8743 for the read-only local dashboard. Setup alone does not produce a meaningful activity history. If the log is empty, follow the capture checks.

A new repository stays local-only without an API key. Re-running setup preserves existing sync settings. Capture scope defaults to all supported sessions for the OS user; use --scope repo to limit it to the selected repository.

Confirm capture works

init installs hooks and arms a 15-minute verification window. Verification pending means setup completed, but runtime delivery has not yet been confirmed. To start a fresh check for Codex in VS Code:

secuarden verify-live --repo . --agent codex --client vscode

Start a new conversation in that repository, ask it to run pwd, and obtain that conversation's exact session ID from the client or its associated local session record. Then check it within the window:

secuarden verify-live --repo . --agent codex --client vscode \
  --session EXACT_SESSION_ID --wait 60s
secuarden status

Replace EXACT_SESSION_ID with the actual ID. For Claude Code, use --agent claude-code and the appropriate intended --client value. A verified result establishes a persisted tool request and matching result; it does not certify tool success, hash integrity, or SaaS delivery. The client label is user-declared, not independently established by hook evidence. See the live-verification guide for activation, exact-session selection, pending results, and stale verification diagnostics.

Platforms and agents

Release build configuration	Architecture
macOS	Apple Silicon (ARM64), Intel (AMD64)
Linux	AMD64

Linux ARM64 and Windows binaries are not currently included in the release configuration. This matrix describes configured builds, not a claim that fresh installation has been validated on every platform.

Claude Code is supported. Codex support depends on runtime hook capabilities and required trust. Cursor, Copilot, and Windsurf integrations are not included.

Demo

See repository-scoped setup, inspect captured activity and sensitive-path access, and verify the stored evidence with the current CLI.

Reproducible walkthrough using synthetic Claude Code hook payloads and actual CLI output. Runs in a disposable local environment; no live agent or SaaS call. The reported test failure is simulated. Hash verification checks stored records, not capture completeness.

Read the demo transcript · Reproduce the demo

What you can inspect

• Agent activity: captured file reads and writes, commands and reported outcomes, MCP calls, sensitive-path access, and session identity.
• Local summaries: activity, agent, timeline, file, command, and attention views with drilldowns. Unknown command outcomes remain distinct from success.
• Evidence integrity: per-session hash-chain verification and AI Bill of Materials (AI-BOM) export.
• Committed changes: Git-anchored change records, optional Codex observation coverage, and signed change receipts.

secuarden report --since 24h --md
secuarden ledger build --repo . --head HEAD --base-ref origin/main

The ledger example requires a locally available origin/main ref. It records committed changes; uncommitted work is excluded. See activity commands and ledger workflows for the full instructions.

Privacy and optional sync

Data	Default location
Captured activity	~/.secuarden/secuarden.db
Git ledger and signed receipts	~/.secuarden/ledger.db
Independent Codex observation inbox	~/.secuarden/observations
Repository integration settings	~/.secuarden/config.json

Activity capture applies sensitive-path detection and secret-redaction patterns before storage. Sensitive paths may still appear as evidence. Redaction reduces exposure; captured metadata should still be treated as sensitive.

Optional session sync sends repository/session metadata and redacted event fields according to the upload contract. API keys and server URLs are configured per repository, outside the worktree:

secuarden init --repo . --api-key YOUR_API_KEY

Omitting --api-key on a later setup run does not disable existing sync. Signed receipt creation separately requires SaaS key registration and attempts to upload its signed evidence envelope. Independent observation capture is opt-in. See repository configuration and receipts.

What the evidence establishes

Secuarden records what supported hooks deliver. Missing hooks, unsupported runtimes, or incomplete sessions can leave gaps. Hash verification detects changes to covered records; it does not establish complete capture or prove who wrote every line. Observation correlations report inferred links and limitations.

Parallel tool calls can have complete request/result pairs while repository checkpoints remain partial because their before/after snapshots overlap. Live verification and checkpoint completeness are separate checks; a successful round-trip does not remove checkpoint uncertainty.

Secuarden complements code and security scanners. It does not replace SAST, secret scanning, dependency analysis, prompt-injection testing, or organizational approval controls. Its CI gate evaluates captured evidence against deterministic risk thresholds. Capture does not direct the agent's edits; synchronous hooks and network operations can add latency.

Verify, export, and gate

secuarden verify
secuarden bom --since 30d > secuarden-ai-bom.json
secuarden gate --since 24h --fail-on high

Run these where the relevant captured database is available. A clean CI runner does not automatically have evidence from a developer's machine. Legacy records without complete hashing cannot gain integrity coverage retroactively; see verification behavior.

Guides

• Setup, agent configuration, capture checks, and uninstall
• Live capture verification and troubleshooting
• Activity commands, dashboard, reports, and verification
• Git ledger, Codex observations, and signed receipts
• Architecture and evidence contracts

Contributing

Documentation fixes, synthetic hook fixtures, redaction improvements, and agent integrations are welcome. Open an issue before starting significant work.

Building requires Go 1.25 or newer and Git. From this checkout:

make build
make test
./bin/secuarden --help

Use a stable installed binary for setup: make install copies it to /usr/local/bin/secuarden and requires write access there. Avoid go run for setup or hook installation because hooks reference the executable's path.

License

Apache 2.0. Built by Secuarden.