phinq

The runtime checkpoint for AI agents. Phinq watches every action your agent or automation takes, pauses the risky ones until you approve (terminal, Telegram, or Slack), and keeps a tamper-evident record of everything — verifiable evidence of human oversight.

Two-minute start

npx @phinq/phinq

The setup wizard detects what you run (Claude Code, Codex, Gemini CLI, Hermes, MCP servers…), asks three plain-English questions, and prints the one line to paste. Then:

phinq start     # start the checkpoint
phinq watch     # live view of anything held

Your existing API key keeps working — Phinq passes it through and never stores or logs it. It starts in watch-only mode: nothing is blocked until you say so.

What you get

• Every action classified by risk — deterministic rules, not another LLM. Deletes, credential access, payments, bulk sends, permission changes, runaway token burn.
• Approve/Deny from wherever you are — terminal, Telegram, or Slack. No answer in time = auto-deny.
• A record you can prove — hash-chained audit log; phinq report turns it into a human-oversight report (false-hold rate, damage prevented, token spend).
• A checkpoint that learns — phinq learn turns your approve/deny history into cited policy proposals.
• MCP gateway — wrap any MCP server: npm i -g @phinq/phinq then phinq-mcp --enforce -- <server command>.
• Works with everything — four wire dialects (OpenAI Chat Completions incl. streaming, OpenAI Responses, Anthropic Messages, Gemini generateContent) cover Claude Code, Codex, Gemini CLI, Hermes, LangChain, CrewAI, Mastra, and any agent with a base URL. No base URL? Gate any action from any language: POST /phinq/gate {"name": "...", "arguments": {...}} → {"allowed": true/false} — holds buzz your phone first.

Everything else

Full docs, configuration, the risk model, SDKs (TypeScript in-process + pip install phinq for Python), and the LiteLLM integration: phinq.co/docs · github.com/phinq-co/phinq

MIT licensed.