KitForge – generate an AI agent manifest; scaffold with enforced approval gates
KitForge is an open-source framework that enforces AI agent safety controls through deterministic gates, including authority budgets, tamper-proof audit trails, human-in-the-loop approval, and circuit breakers. It offers a browser-based design tool and Python CLI to generate agent code with enforced safety measures.
Why this exists. Most “agent safety” frameworks describe guardrails the code never enforces — an audit verify that returns true, budget counters that never block. KitForge is built the opposite way: the controls are deterministic gates around the model, and the repo ships the tests that prove they stop a real violation. Run python -m pytest tests/ and watch a tampered audit log fail verification and a rejected approval halt the action.
What enforces
Authority budgets that block
Tool-call, token, and wall-clock counters decrement and raise when exhausted. The agent loop stops — it does not log a warning and continue.
Audit trail that fails on tamper
An append-only, HMAC-chained log. verify() recomputes the chain and returns false if any entry was altered, reordered, or deleted. Not a no-op.
Human-in-the-loop that halts
A rejected or timed-out approval raises and the guarded action never runs. Approval is required structurally, not requested politely.
Circuit breaker + output validation
A breaker opens after repeated tool failures and refuses further calls; tool output is validated against its declared schema before it touches agent state.
Download KitForge (.zip)Python 3.10+ · MIT · ~30 KB
Design one in your browser
Sketch an agent here. You get a valid blueprint.json to save and feed to KitForge locally — and a one-click check of the design in the Compliance Scanner. The code itself is generated on your machine by the Python CLI; this builds the blueprint it runs on.
Agent name
What it does
Tools tick the ones that should need human approval
approval
Tool-call budget
Audit level
Add a name and at least one tool to build.
Quick start
unzip kitforge.zip && cd kitforge pip install pydantic # the only runtime dep python -m pytest tests/ -q # 13 enforcement tests pass
generate the demo agent (Quarterly Earnings Analyzer)
python -m kitforge demo -o ./my-agent cd my-agent && export KITFORGE_AUDIT_KEY=$(openssl rand -hex 32) python main.py # runs the loop; HITL gate stops 'publish' until you approve
Honest scope
Python + LangGraph patterns only. The generated graph runner is intentionally minimal; swap in langgraph.StateGraph for production — the safety wiring is identical.
Generated tools are [MOCK] stubs. Wiring them to your real systems (and your real approver — Slack, email, web) is the last mile, and it's yours.
The framework enforces the controls it ships; it is not a substitute for a security review of your specific deployment.
Part of the AgentKits governance tools. See also the Compliance Scanner and Risk Assessment.