AI and agents
Agents held to hard limits, and tools to check them.
Eight repos where an AI agent or model is held to a hard limit: writes and destructive actions wait for a person, answers need evidence, lab instruments stay inside signed-off limits, and model output must match a schema. Two of them check agents themselves, by replaying their tool calls and by auditing MCP servers.
- Repos
- 8
- Passing tests
- 864, one dot each, run in October 2026 the way each repo's CI runs them
- Languages
- Python, Go, TypeScript
The eight repos
Open a repo for its project page: what it does stage by stage, with real runs.
-
triage-graph
134 passing tests
Featured Incident triage with a human approval gate
Multi-agent incident triage on LangGraph with a human approval gate enforced in code, plus a CrewAI port and a written comparison of the two.
- Language
- Python
- Stack
- LangGraph, CrewAI, Pydantic
- Scope
- Runs on fixture data.
- Counted with
uv run pytest(134 passed, 1 skipped)
-
sqlguard-mcp
139 passing tests
Featured SQL for agents; every write waits for a person
An MCP server that lets an agent query a SQLite database and refuses any write until a human approves that exact statement.
- Language
- Go
- Stack
- MCP Go SDK, SQLite, Helm
- Approval
- A human runs
sqlguard approve <token>out of band. Each approval is single-use and bound to the exact statement. SQLite only. - Counted with
go test -race -count=1 -json ./...(65 tests plus 74 subtests)
-
bench-mcp
140 passing tests
Lab instruments for agents, behind interlocks
An MCP server that lets an AI agent drive lab instruments over SCPI, with interlocks from a signed-off limits file that the agent cannot override.
- Language
- Python
- Stack
- MCP Python SDK, SCPI over TCP, Hypothesis
- Counted with
pytest -q
-
agent-replay
78 passing tests
Replays an agent's tool calls on a new model
Records an AI agent's tool calls, replays them against a new model or prompt, and fails the check when the agent starts making a destructive call it never made before.
- Language
- Python
- Stack
- Python standard library only (optional Anthropic and MCP SDKs), pytest, Hypothesis
- Counted with
pytest -q(with the optional Anthropic and MCP SDKs installed, as in CI)
-
mcp-audit
65 passing tests
Audits MCP servers in a traced sandbox
An auditor for MCP servers that connects as a real MCP client, scans what each server tells the model, and calls every tool in a traced, offline sandbox to check it does what its annotations claim.
- Language
- Python
- Stack
- Official MCP Python SDK, Docker, strace
- Counted with
make images && pytest -q(4 of them need the Docker sandbox images, as in CI's sandbox job)
-
rag-knowledge-agent
109 passing tests
Cited answers, or a refusal
A command-line retrieval agent that answers from a folder of documents with citations, and refuses without calling the model when no evidence clears its confidence floor.
- Language
- Python
- Stack
- Claude, ChromaDB, Pydantic
- Counted with
pytest
-
llm-contract-router
105 passing tests
Schema-checked routing across LLM providers
Routes requests across LLM providers behind one interface, validates each answer against your schema, and re-prompts with the constraint that failed.
- Language
- TypeScript
- Stack
- Node.js, Anthropic, OpenAI, zero runtime deps
- Counted with
npm run check
-
local-network-mcp
94 passing tests
Network and SSH tools, denied until opted in
An MCP server for local network, system and SSH tasks, where shell commands, remote commands and process kills stay denied until the operator opts in.
- Language
- Python
- Stack
- MCP, Paramiko, psutil
- Gated
- 4 of its 18 tools: shell, SSH connect and execute, and process kill.
- Counted with
pytest