Spend and safety controls for Claude Code and Codex, running on your machine.
Once a session passes your sub-agent limit, the next launch waits for your yes. Every decision is an Ed25519-signed receipt in a local hash chain. The band above your prompt shows sub-agents against the cap, tokens, the share that went to sub-agents, your plan limit, and how far you are from Claude Code's auto-compaction point.
claude plugin marketplace add agentguard-run/agentguard-codex-plugin
claude plugin install agentguard@agentguard
Free on one machine. No account, no key, nothing uploaded.
npx agentguard-burn why
npx agentguard-burn week
npx agentguard-burn compact-replay
Burn reads your local Claude Code and Codex transcripts and prints what a session spent, what last week's waste causes were, and which /autocompact setting would have saved tokens over your last 7 days, measured on your own sessions with its error band.
- agentguard-codex-plugin: the plugin for Claude Code and Codex
- agentguard-mcp: the remote MCP server (spend caps, signed receipts, setup checks)
- agentguard-skills: skills for Claude Code
- agentguard-hermes: AgentGuard for Hermes Agent
- agentguard-cb: open-source evidence SDK for Stripe disputes
Site: https://agentguard.run. Privacy: https://agentguard.run/legal/privacy. AgentGuard is a registered trademark of Dunecrest Ventures Inc.