Doberman★ FinalistDeep-reviewed
by Tian Herng Tan
What they're building
Doberman - a personalized AI agent security layer with adaptive authorization & runtime guardrails for AI coding agents
AI code reviewrepo: real
Substantial, well-engineered Python codebase (v0.11.0): a transparent MCP proxy that intercepts every tool call and returns PASS/AUTH/BLOCK, with a decision engine, objective rule guardrails (paths/commands/secrets/destinations/smuggled-token channels), a subjective anomaly/homoglyph layer, TOTP step-up auth, roles/boundaries, audit log, and import-linter architectural contracts plus an end-to-end integration test. Public demo video exists ('Doberman - a zero-trust firewall for your AI coding agent'). However it uses ZERO sponsor stack -- pyproject deps are only pydantic/mcp/aiosqlite/pyotp/pyyaml/typer, and the README/CI explicitly enforce 'no proprietary dependency, ever', so Composio/Nebius/Tavily are entirely absent (hard floor on sponsor usage). It is also arguably a guardrail/proxy rather than an autonomous AI agent itself. ⚑ No sponsor integration at all (CI-enforced, by design) -> sponsor_usage floored to 1. Verified via repo + README + public demo title rather than a live hosted UI. Borderline fit for an 'AI agent' build (it governs agents).