I built Kotro — a local control plane that sits between coding agents (Claude Code, Continue, Cline, Cursor via HTTPS bridge) and both the model and MCP tools they use.
One ~15MB Rust binary on localhost. No SaaS required for the sidecar.
What it does
-
MCP plane (
mcp-wrap): pin tool schemas, catch rug-pulls, validate tools/call args, optional signed TaskEnvelope approvals -
LLM plane: scan for prompt injection in tool results (warn by default;
KOTRO_INJECTION_BLOCK=true→ HTTP 400), redaction, cache, budget, kill switch - One dial:
KOTRO_MODE=disabled|audit|enforce - Escape Lab: CI adversarial corpus with documented gaps (not "100% blocked") Honest limit: it is not an OS egress firewall. Agents that shell out or open raw sockets never hit the proxy. ## Try it (no API key)
git clone https://github.com/kotro-labs/kotro-proxy-engine
cd kotro-proxy-engine && make demo-injection
Needs Rust + Go on first build. Dashboard: http://127.0.0.1:9090/dashboard
Install
curl -sL https://raw.githubusercontent.com/kotro-labs/kotro-proxy-engine/main/scripts/install.sh | bash
# or: brew install kotro-labs/tap/kotro-proxy
Repo: https://github.com/kotro-labs/kotro-proxy-engine (MIT)
Feedback welcome — especially whether audit→enforce is the right default path.
Top comments (0)