Skip to content
Agentic Control Plane

Control Claude Code, Codex and Cursor permissions in one command (no signup, runs local)

David Crowe David Crowe · · Updated · 6 min read
claude-code codex cursor tool-policy local
Share X HN LinkedIn

Just want to govern Codex? If you're here to actually see, control, and price every Codex tool call, that's one command — hooks for Bash plus an MCP connector for everything else, in a single installer:

macOS · Linux · WSL
curl -sf https://agenticcontrolplane.com/install.sh | bash
Windows · PowerShell
irm https://agenticcontrolplane.com/install.ps1 | iex

Full Codex install guide →  ·  see your first governed call →  ·  free up to 5 agents

Since this recording, a force-push is held for your approval rather than blocked.

The recording above is real: the live installer, the real CLIs, the real hook path. It runs about a minute and tells one story. The command:

curl -sf https://agenticcontrolplane.com/install.sh | bash -s -- --local

No account, and nothing leaves your machine. Then a Claude Code session is asked to squash the repo and force-push main. The push is stopped at the hook, with the reason shown, before git runs. A curl to an external API has to ask first. The tests run without a prompt, because normal work shouldn’t cost you a click. At the end, tail ~/.acp/audit.jsonl shows every call the agent made, with the decision and the reason on each line.

One update since the recording: the installer now treats a force-push as a hold that waits for your approval, not an outright block. Codex can’t show a hook’s ask as a prompt, so the coverage page is the place to check what happens there.

What the command installs

Three files in ~/.acp, plus a hook entry in each client it finds:

  • decide.mjs is the decision engine. It’s a few hundred lines with no dependencies, and it makes no network calls. It labels a call (Bash.git.push, Bash.curl.api.stripe.com, Write), applies the floor, then walks your rules from most specific to least. Read it before you let it near your shell: source.
  • policy.json holds your rules. Each is allow, ask or deny, set per tool, per git subcommand or per curl host. "Bash.git.push": "ask" is the whole syntax.
  • audit.jsonl is an append-only log of every call: the tool, how it was labeled, the decision, where the decision came from, and the reason.

The safety floor sits under all of it. It always blocks rm -rf on root or home, mkfs, dd to a disk device and fork bombs. It holds force-pushes for your approval. It parses the command, so rm -r -f ~/ and bash -c "rm -rf ~" after an && are caught as the same command in different clothes. The bypass write-up is why plain string matching wasn’t enough.

What it isn’t

It isn’t isolation. A prompt-injected agent can still read what it can read, and a devcontainer is still the right call for untrusted input. This controls actions and a container limits the process, so they stack.

The audit log is a plain file on the same machine. It’s a record, not tamper-proof evidence.

If you never use --dangerously-skip-permissions, Claude Code’s own prompts already do a lot of the work. The ranked comparison lays out the options.

What this adds is the cross-tool part: one file, one floor and one log for Claude Code, Codex and Cursor, instead of three config formats that drift apart. How much each tool lets the hook do differs, and the coverage page lists it.

It’s MIT, and every file the installer writes is documented. If you outgrow one machine, the same installer without --local connects a workspace, with shared policy, priced sessions and one console across everyone’s agents.

Does this need an account or send data anywhere?

No. In –local mode there is no signup, no token, and no network call to make a decision. The policy engine, the policy file and the audit log are plain files in ~/.acp on your machine.

What does the safety floor do?

It always blocks recursive force-deletes aimed at root or home, mkfs, dd onto a raw disk device, fork bombs and chmod -R 777 /. A force-push is held for your approval, not blocked. The floor parses the command, so reordered flags and bash -c wrappers don’t get past it, and it applies whatever your policy file says.

Which agents does it cover?

Claude Code, OpenAI Codex CLI and Cursor. The installer detects what you have and wires the same hook into each, and one policy file applies to all of them. Codex needs a one-time /hooks trust review inside Codex, and the installer tells you. Cursor’s and Codex’s ask support is limited, so check the coverage page for how a held call reaches you.

Is this a sandbox?

No. It controls actions in the call path. A container limits what the process can reach. If your agent handles untrusted input, run a devcontainer too. They stack.

What's the difference between local mode and the cloud?

Local mode is the free individual runtime: decisions on your machine, one policy file, one audit log. Re-running the installer without –local connects a workspace, which adds shared policy for a team, priced sessions and one console across everyone’s agents.

Share X HN LinkedIn
Get the next data drop
What agents actually cost, new tool-surface captures, and the occasional incident post-mortem — sent when we publish something worth your inbox, not on a schedule. Unsubscribe anytime.
Share: Twitter LinkedIn
Related posts

← back to blog