Local execution control · Free under Apache-2.0
Let agents work.
Keep control of their actions.
Put a permission check in the code that runs the action. Decide what may execute, require operator approval when needed, and keep evidence of the decision.
Current stable release: 2.0.0. Fixed run deadlines, optional supervisor-only liveness, private operator state and readable denials. Conflicting recognized destinations and damaged-store startup are refused. Stable MCP bundle and editor setup. Published qualification is linked separately from this historical replay; prior RC3 GUI/model tests are not certification. We will review feedback weekly, with 2.0.0 as the baseline through October 19. Tested Apache-2.0 contributions are welcome.
Inspect a real execution check
An allow runs the body. A denial leaves it uncalled.
This browser replays recorded results from a real disposable-file evaluation on Apple Silicon macOS, using the installed 1.1.0-rc.1 Python client and native binary. These are historical recorded results, not a rerun of stable 2.0.0. No agent or sidecar runs in this Space.
Read the complete recorded report
Nine synthetic checks; zero model requests. Automated fixture approvals are for testing only. Your production operator must review approvals, and your own protected routes need acceptance tests. Download the report.
Build it in
Embed Rust at your executor's dispatch point, or put a local HTTP sidecar around the real Python or Node callback.
Bolt it on
Wrap a trusted stdio MCP server and gate its routed tool calls. Keep operator credentials outside model-controlled code.
Try it without an AI account
Download the native archive for your OS and verify its checksum. From the extracted directory on macOS or Linux:
chmod 0755 deadbolt
./deadbolt drill
python3 examples/evaluate.py --binary ./deadboltExpected: deadbolt drill ok, then a JSON report with "passed": true. Python 3.9+ is needed for the evaluation. Windows commands and setup are in the guide.
Stable downloads and qualification · Historical 1.0.3 release · Installation guide
The Python wheel does not bundle the executable; install the matching engine separately. Python client 2.0.0 on PyPI · Rust CLI/library from exact stable source. Read the migration guide before upgrading. No npm package is offered here.
Watch it work · 4:31 overview
Visible English captions throughout. You can follow on mute.
Overview on YouTube · Full developer walkthrough · 21:17 · Read the transcript
Videos were recorded before prerelease publication; candidate-status narration reflects that date. Cedar narration is AI-generated. Harness uses a deterministic local demo provider; its dispatcher and file effects are real. Presentation slides explain the system.
Make the boundary explicit
DeadBolt protects actions routed through its gate. Your trusted executor supplies identity, tools, destinations and reported spend. The operator controls policy, approvals and revocation. Ordinary Rust/HTTP admission leaves unset tool policy open; MCP admission requires a nonempty tool allowlist. Configure policy before dispatch.
It does not sandbox arbitrary code, cancel an already running effect, independently meter billing, or make an external effect transactional with admission. Evidence records are not a compliance certification.
Read the trust boundary · Run an integration pilot · Share an integration problem