How to govern AI agent tool calls with Haldir

[ExposureGuard]

Haldir sits between your AI agent and the tools it uses. Every MCP tool call is intercepted, authorized, and logged.

Quick start:

pip install haldir

The problem: AI agents call APIs, spend money, and access credentials with zero oversight.

The fix: Haldir's proxy mode intercepts every tool call:

1. Session validated (does this agent have a valid session?)
2. Permissions checked (can this agent call this tool?)
3. Spend limits enforced (has the budget been exceeded?)
4. Policies applied (is this tool on the allow list?)
5. Approval checked (does this need human sign-off?)
6. Action logged (immutable audit trail)
7. Call forwarded to upstream MCP server

Endpoints: haldir.xyz/docs
OpenAPI: haldir.xyz/openapi.json
Smithery: smithery.ai/server/haldir/haldir (98/100)