About
Agents move fast. Strazh holds them accountable.
Strazh is a local safety gate for AI coding agents, built by Graphenlabs.
Why Strazh exists
Coding agents are useful precisely because they act on their own: they read a prompt, decide what to touch, and run. That same autonomy is the risk: an agent that can write to your filesystem and execute processes can also write to the wrong repository, read a secret it should never have seen, or run away past a timeout. Most of the tools built for coding agents focus on making them better at writing code. Very few focus on the moment before that: deciding whether the agent should have been allowed to run at all.
Strazh is built around one idea: the same program that launches your coding agent should be the one thing that can stop it. Not a policy documented somewhere, not a convention your team agrees to follow: an actual process boundary that sits between the command and the process, every time.
What it actually does
Four mechanisms, all enforced locally, none of them optional:
- Allow-list. An agent can only dispatch into working directories you’ve explicitly registered. Anything else is rejected before a process ever spawns.
- Environment scrub. Secrets sitting in your shell environment never enter the dispatched agent’s process.
- Tree-kill. If a dispatch runs past its timeout, the whole process tree ends in one kill: not just the parent, not a process left orphaned.
- Audit log. Every dispatch, approved or rejected, is written to a local, queryable record. The ones that never ran are on the record too.
See it happen against a real dispatch, including a rejection, on the homepage, or read the exact commands in Install & setup.
Where the code stays yours
Strazh decides what runs. What lands in your repository is a separate concern, and conflating the two would mean trusting one tool with both the process boundary and the code-review judgment call. Coval, our diff-planning and diff-review layer, handles that half: plan the blast radius before an agent touches anything, review the diff like a colleague wrote it, approve or reject on your own call.
Local by design
Strazh is a local-first desktop application. Your dispatch requests, the repositories you target, and your audit log live on your own machine, not on a server we operate. There’s no dashboard watching your dispatches from the outside, because there’s nothing to watch remotely in the first place. The full detail of what does and doesn’t leave your machine (including the two narrow, opt-in exceptions for cloud-hosted agents like Devin) is in our Privacy Policy.
Graphenlabs
Strazh is built by Graphenlabs. We’re a small team building safety infrastructure for the AI-agent era: Strazh is our first product, distributed as a local app and an MCP server, gating every major coding agent (Claude Code, Codex, Cursor, and Devin) behind the same enforcement point.