Make your agent's worst day survivable.
YOLO mode, on belay. Set limits on the agent instead of relying on the
model's judgment about what's safe. Works with Claude Code, Pi, and Grok Build.
Why · Install · What it blocks · Per-session · Config · Limitations
# Skip every permission prompt. It still can't reach the network or leave this repo.
BELAY_NETWORK_MODE=sandbox \
BELAY_WORKSPACE_GUARD=on \
BELAY_ALLOWED_ROOTS="$HOME/Github/my-app" \
BELAY_SANDBOX_DENY_WRITE="$HOME/Github" \
BELAY_SANDBOX_ALLOW_WRITE="$HOME/Github/my-app" \
claude -p "fix the scroll bug" --dangerously-skip-permissionsBuilt by Derek Larson. Backstory: Keys to the Castle.
I wanted Claude Code running without my involvement, next to sensitive data. Claude Code has no way to say "you can touch this folder, but nothing else on the machine."
Why not a container or a worktree? Those isolate the repo, not your credentials — you mount those in yourself, because the work needs a git identity and an npm token. Belay draws the line the other way: the agent keeps your machine and your tools, and loses the parts it has no business in.
/plugin marketplace add derek-larson14/belay
/plugin install belay@belay
/belay:setup
git clone https://github.com/derek-larson14/belay && cd belay
./belay install
./belay status./belay install pi # or: claude | grok | all
./belay off pi| Harness | Wires into |
|---|---|
| Claude Code | ~/.claude/settings.json |
| Pi | pi install → ~/.pi/agent/settings.json |
| Grok Build | ~/.grok/hooks/belay.json |
belay off # unwire every harness — scripts and config stay put
belay off claude # or one at a time: claude | pi | grokbelay on puts it back. To remove it entirely:
belay uninstall # unwire everything, drop the deny rules we added
belay uninstall --purge # also delete ~/.belay (config + audit log)Claude Code plugin users: also /plugin uninstall belay@belay.
on, off, and uninstall only touch entries Belay wrote. Other hooks in your settings.json stay where they are, including hooks that share a matcher with ours. uninstall removes exactly the permissions.deny rules /belay:setup added — recorded at install time — and leaves your own rules alone.
Every tool call goes through four guards. First deny wins.
Path guard — blocks sensitive paths (credentials, browser sessions, keychains, clipboard, shell history, …). Categories toggle in config.
Write guard — blocks writes to shell rc files, LaunchAgents, SSH authorized_keys, and similar.
Network guard — sandbox (macOS: no network for shell commands), pattern (blocks known exfil patterns), or off.
Workspace guard — optional. File tools only see the project dirs you list.
Catches the usual spellings of sensitive paths, for example:
~/.ssh/id_rsa$HOME/.aws/credentialscat ~/Library/Messages/chat.db
Pick folders the agent may not write to. Optional allow-list for exceptions.
BELAY_SANDBOX_DENY_WRITE="$HOME/Github/exec" \
BELAY_SANDBOX_ALLOW_WRITE="$HOME/Github/exec/scratch/build" \
claude -p "build the feature" --dangerously-skip-permissionsEnv vars beat config for that run only.
BELAY_NETWORK_GUARD=on \
BELAY_NETWORK_MODE=sandbox \
BELAY_WORKSPACE_GUARD=on \
BELAY_ALLOWED_ROOTS="$HOME/Github/my-app:$HOME/Github/my-lib" \
claude -p "fix the scroll bug" --dangerously-skip-permissionsBELAY_NETWORK_GUARD=on
BELAY_PATH_GUARD=off
BELAY_NETWORK_MODE=sandbox
BELAY_ALLOWED_ROOTS="/a:/b"
BELAY_SANDBOX_DENY_WRITE="/protected/path"
BELAY_SANDBOX_ALLOW_WRITE="/protected/path/ok"
# Path guard categories
BELAY_PATH_CAT_CREDENTIALS=off
BELAY_PATH_CAT_CLIPBOARD=off
BELAY_PATH_CAT_BROWSER_SESSIONS=off
# Also: messages, keychains, password-managers,
# system-data, shell-history, claude-internals, browser-hijacking
# .env dumps (Read/cat/grep). Processes may always load them.
BELAY_ENV_FILES=project # block | project | allowbelay install [claude|pi|grok|all]
belay status
belay on / off [harness|all]
belay uninstall [--purge]
belay log [N]
belay test
Claude Code slash commands:
/belay:setup
/belay:scan
/belay:configure
/belay:toggle
Credentials — SSH keys, AWS creds, API tokens, Docker/Kubernetes config
.env files — dumping contents into the model (Read, cat, grep). Processes may load them (source .env && npm test, python app.py). block (default) denies every dump; project allows project .env dumps and still blocks $HOME/.env; allow turns the check off. Templates (.env.example) are never blocked.
Browser sessions — Cookies and local storage for Chrome, Arc, Firefox, Safari, Brave, Edge, Dia
Password managers — 1Password vaults, system keychains, GNOME keyring, GPG keys
Messages and email — iMessage, Mail, Signal databases
Clipboard — pbpaste, pbcopy, xclip, xsel
Shell history — .bash_history, .zsh_history, .psql_history, .python_history
Network — no network for shell commands on macOS sandbox mode; patterns for cookie theft, reverse shells, scp/rsync
Writes — folders you mark off-limits; also persistence targets (LaunchAgents, crontab, shell rc, SSH authorized_keys)
Browser hijacking — --remote-debugging-port, Puppeteer/Playwright connect, Chrome DevTools Protocol
Highest wins, per key:
BELAY_* env > <project>/.belay.toml > ~/.belay/config.toml > defaults
A project file that sets one key only overrides that key.
With no config you get path + write guards on; network and workspace are opt-in. belay status shows every value and which layer set it.
belay.example.toml is the reference. belay install copies it to ~/.belay/config.toml once and never overwrites it.
[path-guard]
enabled = true
env_files = "block" # "block" | "project" | "allow"
[path-guard.categories]
credentials = true
browser-sessions = true
messages = true
keychains = true
password-managers = true
system-data = true
shell-history = true
claude-internals = true
clipboard = true
browser-hijacking = true
[write-guard]
enabled = true
[network-guard]
enabled = false
mode = "pattern" # "sandbox" | "pattern" | "off"
[workspace-guard]
enabled = false
allowed_roots = ""
[audit-log]
enabled = true
path = "~/.belay/audit.jsonl"Project override: .belay.toml at the project root.
Every tool call goes to ~/.belay/audit.jsonl: what was touched, when, and whether a guard blocked it.
Blocks the agent from editing Belay itself, its config, or the harness hook files.
BELAY_SELF_PROTECT=on claudeString matching is imperfect. On macOS, sandbox mode is the strong path.
Workspace guard covers file tools (Read/Write/Edit/Grep/Glob), not the shell. Use BELAY_SANDBOX_DENY_WRITE if you need write limits on shell commands.
sandbox-exec is macOS only. Linux uses pattern mode for network and has no write sandbox.
| Platform | Status |
|---|---|
| macOS | Full |
| Linux | Pattern network guard, no write sandbox. Everything else works. |
cd /path/to/belay
./belay testRuns the guard suite, normalize/dispatch, and the Pi bridge.
- Guard the Monitor tool. It runs shell commands like Bash, but the hook never fired for it, so with the network sandbox on,
curlthrough Monitor still reached the internet. Monitor now goes through every Bash check and the same sandbox. Existing installs: rerunsetup.sh(or/belay:toggleoff and on) so the hook matcher picks upMonitor. - The sandbox wrap now keeps the rest of the tool input. It used to replace the whole input with just the command, which dropped Bash
timeoutandrun_in_background.
- Fix false credential blocks caused by regex text expanding into dotfile names during Bash tokenization. Actual unquoted dotenv dump globs remain blocked.
- Resolve relative workspace paths against the session cwd and allow new files inside the workspace. Match directory boundaries so a sibling such as
project-oldis excluded. - Resolve symlink destinations for new paths, including dangling links, and deny loops or unresolved paths. Check the session cwd for Grep/Glob calls with no explicit path.
- Add regression coverage for regex/glob handling, new paths, workspace boundaries, implicit search paths, and internal/external symlink destinations.
Issues and PRs welcome.
MIT
