0
0
via GitHub · Posted Jul 17, 2026 · 1 min read

Gensee Crate: AI Agent Runtime Safety

GenseeAI/gensee-crate
Tool

Runtime safety for AI coding agents with real-time enforcement, system-event monitoring, and long-horizon provenance. Supports Claude Code, Codex, Antigravity, Copilot, Omnigent on native macOS and Linux.

123Stars
12Forks
1Open issues
1Watching
Rust Apache-2.0 v0.3.3 Updated 4 days ago

At a glance

Rust Actively maintained Commercial OK

Gensee Crate provides real-time safety enforcement and monitoring for AI coding agents through policy enforcement, system-event tracking, and provenance lineage. It integrates with Claude Code, Codex, Copilot, and other agents on macOS and Linux to capture and control file operations, command execution, network access, and other risky behaviors.

0 comments

README


Why Gensee Crate

  • Keep attention on the work that needs you. The Review Queue groups supported background activity with its originating request and surfaces scope drift, blocked actions, approval requests, and failed verification. Routine activity stays in history.
  • Set boundaries before risky actions. Local policy can allow, ask, or block supported tool actions and protect sensitive paths. Enforcement depends on the integration and execution environment.
  • Reduce repeated interruptions. Remember explicitly approved actions within a session or project, or allow matching reads to a chosen file or folder. False-positive feedback is separate from permission: it does not silently grant access or weaken a rule.
  • Review and recover changes. Personal creates recovery points for supported Git-workspace changes. Team provides disposable Linux workspace forks that you can inspect, merge, promote, or discard.
  • Know what was observed. Review action evidence alongside policy decisions. Gensee reports monitoring gaps as its own coverage issues, separately from agent warnings.

Benchmark results

Preliminary AgentCanary results show Gensee Crate improving defense rate across memory-poisoning, long-horizon, and prompt-injection threat types with low runtime overhead.

Preliminary AgentCanary benchmark results

Gensee Crate Personal

What it is

Gensee Crate Personal is a local-first macOS app and CLI for people working with AI agents. Current integrations include Codex, Claude Code, Cursor, GitHub Copilot, Antigravity, and Omnigent, plus a macOS endpoint-visibility pilot for Claude Cowork. Your policy, agent events, reviews, and feedback remain in your local Gensee store.

Coverage varies by integration. Cowork records supported host activity and VM boundaries; guest commands and cloud execution remain outside endpoint coverage. See the Cowork integration guide for setup and limits.

What it adds

  • A Review Queue that groups work by request and shows what needs attention.
  • Scope-drift detection that compares declared tool intent with file mutations independently observed by macOS Endpoint Security, while keeping routine workspace and temporary-file activity quiet.
  • Smart recovery points before risky Git-workspace changes, with restore actions in the review.
  • Configuration audit for instructions, skills, MCP servers, hooks, permissions, plugins, command rules, and other inputs that can change agent behavior.
  • Scoped, revocable approvals for supported repeated actions, plus separate false-positive feedback for triage. See review and approval controls.
  • Local policy enforcement, verification freshness, activity highlights, notifications, a menu-bar summary, and visible sensor health.

Download the macOS app

⬇️ Download Gensee Crate Personal for macOS

The signed app bundles the Gensee backend and SQLite support. It does not require Homebrew, Rust, Xcode, jq, or a separate SQLite installation. See the Gensee Crate Personal for macOS guide for installation, first-run setup, Apple approvals, harness protection, and local troubleshooting.

See it in use

Use the CLI instead

Install the CLI and initialize the local store:

curl -fsSL https://raw.githubusercontent.com/GenseeAI/gensee-crate/main/scripts/install_oss.sh | bash
export GENSEE_HOME="${GENSEE_HOME:-$HOME/.gensee}"

Configure the harnesses you use:

gensee setup codex --gensee-home "$GENSEE_HOME"
gensee setup claude-code --gensee-home "$GENSEE_HOME"

Run an agent through Gensee when you want an explicit managed session:

gensee run -- codex
# or
gensee run -- claude

Inspect the results:

gensee audit codex
gensee run list --json
gensee timeline --latest
gensee status --json

The desktop app is the recommended macOS experience. The CLI remains useful for automation, terminals, and Linux workstations. See Claude Code hook setup, policy, configuration audit, and run and sandbox modes for the complete command-line workflow.

Gensee Crate Team

What it is

Gensee Crate Team is the self-hosted path for small teams and businesses that want to operate their own Gensee deployment and agent environments. Agents run on a prepared remote Linux host. The team keeps control of its workspace, policy, credentials, runtime, evidence, and lifecycle decisions.

What it adds

The operating principle is simple:

intent
  → capability decision
  → lease, mediator, cell, or workspace fork
  → observed effects
  → merge, promote, or discard
  → revocation and cleanup
  • Transactional workspace forks. tclone creates low-latency, whole-workspace forks for one or several approaches. Each fork can be inspected and tested before a human merges it, promotes it, or discards it.
  • Bounded authority. Request-scoped capability decisions and short-lived leases limit filesystem, network, repository, workload-identity, database, and external-action authority.
  • Credentials stay on the host. The capability broker owns credential material and gives cells opaque lease IDs, scoped handles, or trusted gateway endpoints instead of broad secrets.
  • Independent evidence. Host observation, process lineage, effect manifests, replay plans, promotion receipts, and cleanup journals make it possible to explain what occurred and whether it stayed within the granted authority.
  • Promotion is a policy decision. Manifest violations, incomplete evidence, failed cleanup, expired authority, or missing commit tokens can prevent work from becoming durable.

The strongest end-to-end enforcement today is in tclone capability cells and network mediation. Additional capability backends are under active development; see the roadmap for the current boundary.

Install on a Linux host

Install Gensee Crate:

curl -fsSL https://raw.githubusercontent.com/GenseeAI/gensee-crate/main/scripts/install_oss.sh | bash
export GENSEE_HOME="${GENSEE_HOME:-$HOME/.gensee}"

Then prepare the remote host with the tclone-enabled os4agent runtime, rootful Podman with btrfs, and a tclone image. Follow the tclone host setup rather than copying host-storage settings between machines.

After host preparation, define the wrapper used by the tclone workflow:

export GENSEE_TCLONE_PODMAN="$HOME/os4agent/podman-tfork.sh"
export GENSEE_TCLONE_IMAGE="${GENSEE_TCLONE_IMAGE:-localhost/gensee-tclone-webtop:tmux}"
export GENSEE_TCLONE_AUTHORITY_ROOT="${GENSEE_TCLONE_AUTHORITY_ROOT:-/var/lib/gensee-boundary}"
export GENSEE_TMP_ROOT="${GENSEE_TMP_ROOT:-/tmp}"
export TMPDIR="$GENSEE_TMP_ROOT"
sudo install -d -o root -g root -m 0700 "$GENSEE_TCLONE_AUTHORITY_ROOT"
# Optional: set this when os4agent uses a dedicated btrfs rootful Podman store.
# export CONTAINERS_STORAGE_CONF="$GENSEE_HOME/tclone-btrfs-storage.conf"

alias gensee-tclone='sudo env \
  PATH="$PATH" HOME="$HOME" TERM="${TERM:-}" TMUX="${TMUX:-}" \
  GENSEE_HOME="$GENSEE_HOME" \
  GENSEE_TCLONE_PODMAN="$GENSEE_TCLONE_PODMAN" \
  GENSEE_TCLONE_IMAGE="$GENSEE_TCLONE_IMAGE" \
  GENSEE_TCLONE_AUTHORITY_ROOT="$GENSEE_TCLONE_AUTHORITY_ROOT" \
  CONTAINERS_STORAGE_CONF="${CONTAINERS_STORAGE_CONF:-}" \
  GENSEE_TMP_ROOT="$GENSEE_TMP_ROOT" TMPDIR="$TMPDIR" \
  gensee'

Run and fork agent work

Start the source agent in the prepared runtime:

gensee-tclone run --runtime tclone -- codex

For a passive baseline with no in-container Gensee hooks or control channel, launch Tclone in observe-only mode. Host lifecycle/cgroup attribution remains available for external collectors, and the resulting source cannot be forked:

gensee-tclone run --runtime tclone --observe-only -- codex

Create one fork or compare multiple approaches:

gensee-tclone run list --json

gensee-tclone run fork <source-run-id> \
  --copies 2 \
  --name try-upgrade \
  --approach 'minimal compatible upgrade' \
  --approach 'aggressive latest-version upgrade' \
  --attach tmux:right \
  --json

Examine results and decide what persists

gensee-tclone run summary <fork-id> --json
gensee-tclone run diff <fork-id> --json
gensee-tclone run compare <parallel-fork-id> --json

# After an explicit human decision:
gensee-tclone run choose <parallel-fork-id> --merge
# or: --promote
# or: --discard-all

Use gensee timeline, gensee status --json, and the local Gensee dashboard to examine policy decisions, runtime evidence, effects, cleanup, and promotion outcomes. The tclone guide and capability broker guide describe the complete host, lease, mediation, and lifecycle model.

Open agent-security traces

The security-traces collection publishes redacted, schema-validated experiment corpora for detector development, forensic replay, and capability-policy research.

The collection currently includes two blind autonomous-agent scenarios: an overbroad inference credential that enabled hosted web-search effects, and an authentic, deliberately vulnerable Nexus 2 package-service fixture. The Nexus release records four completed classifier trials; one produced an independently confirmed L4 package-service boundary escape and three completed without an observed escape.

Each release includes its exact task prompt, normalized telemetry, model event stream, ground truth, provenance, redaction and coverage records, checksums, and offline validation/replay/scoring tools. The traces contain no credentials, private holdout labels, raw SCAP, runnable service infrastructure, or executable network replay.

Roadmap

  • Personal: richer verification results, more harness integrations, quieter request-level decisions, and broader independent network evidence.
  • Team: more capability adapters and trusted mediators, a generalized dispatcher across effect domains, stronger remote evidence export, and counterfactual replay before policy changes.
  • Integrations (planned): Glean Agents and n8n workflows, expanding coverage for knowledge work and business automation.
  • Both: keep deterministic policy and evidence portable while reducing the supervision needed for routine work.

Follow the detailed project roadmap and open issues for current work.

Documentation

Gensee Crate is available under the Apache 2.0 license.

Comments (0)

Sign in to join the discussion.

No comments yet

Be the first to share your take.