Integration · Coding agent
Kyvvu for Claude Code
Claude Code's own hooks send every tool call to the engine first: Bash, Read, Write, Edit and the rest. No code changes.
Set up
One command, then every step is checked.
- Every session
- The CLI, the VS Code and JetBrains extensions, and Claude Code in GitHub Actions.
- One call, not the session
- A block denies that tool call. Claude sees why and can take another route.
- Sensible defaults
- The Claude Code safety pack: no secrets sent out after a .env read, no rm -rf /, no force push, stay inside the project.
# Connects to the API, registers the agent, installs the hooks$ kyvvu claude setupHow it works
Claude Code's own hooks, no daemon.
Each hook runs as a short-lived process. Session state, step history and the policy cache are kept in ~/.kyvvu/integrations/claude-code/ between calls.
| Hook | What Kyvvu does | ||
|---|---|---|---|
| SessionStart | Fetches this agent's policies and starts a task. | ||
| PreToolUse | Checks the tool call against the policies and everything the session did before it: allow or deny. | ||
| PostToolUse | Records the completed call in the audit trail. | ||
| SubagentStart, SubagentStop | Links a sub-agent's steps to the step that spawned it. | ||
| SessionEnd | Sends the trace to the platform and ends the task. |
Monitor first
Watch, then enforce.
Setup starts in monitor mode: every call is checked and violations are logged, but nothing is denied. Switch enforcement on when the policies fit.
- Fail open or closed
- Open allows a call when a check fails. Closed denies it, and also when no policies are loaded, even in monitor mode.
- Per project
- kyvvu claude hooks install --project writes the hooks to the project's .claude/settings.json, to commit with it.
- Your own logging
- Send traces to the platform, a file, stdout or an OpenTelemetry collector (OTLP).
$ kyvvu claude statuskyvvu claude status Initialized: yes Agent ID: agent_3f9c2a1b7d4e8f60 Agent key: claude-code-laptop-ada Owner: ada@example.com Environment: development API URL: https://platform.kyvvu.com Log location: https://platform.kyvvu.com Log format: kv Incident location: https://platform.kyvvu.com Incident format: kv API key: *** Mode: monitor (violations are logged, not blocked) Fail mode: open Policies: 20 cached Policies for: agent_3f9c2a1b7d4e8f60 Enforcement: evaluating 20 policies (monitor mode: violations are logged, not blocked) Sessions: 0 active Permissions: private (0600 files, 0700 directories) Hooks: installedDefault policies
Safe defaults, nothing to assign.
When setup registers the agent, the platform applies Kyvvu's baseline in its Claude Code form.
| Manifest | Policies | What it covers | |
|---|---|---|---|
| Claude Code safety | 13 | Credential exfiltration after a secret read, destructive commands (rm -rf /, git push --force, git reset --hard), staying inside the project, runaway exec. | |
| EU AI Act, minimal and limited risk | 7 | Registration documentation and Article 50 transparency. |
Both are in the open policy library: Claude Code safety and EU AI Act, minimal and limited risk. Add more from the dashboard or with kyvvu assign-manifest.
Good to know
- Starts in monitor mode: violations are logged, not blocked. Turn enforcement on with kyvvu claude set enforce true.
- Reading a secret file (.env, .pem, .key) taints the session: after it, the tainted-path policies block Bash and further reads, so nothing read can be sent out. /clear starts a new session.
- macOS and Linux, Python 3.11 or newer. Claude Desktop isn't supported: it uses MCP, not hooks.
- Detecting secret reads and write targets in Bash commands is best effort, not a shell security boundary. Keep OS-level controls alongside it.
Next
Try it before you install it.
In your browser
The playground
Build an agent's path, pick a policy and watch the engine decide.
On your machine
Five minutes, no account
kyvvu try runs a demo agent against the bundled baseline.