Aigenos Flow

At a glance

Aigenos Flow is an open-source Python library that turns the risky parts of an LLM agent, like payments, refunds or bookings, into fixed steps written in a SKILL.md file. The model only turns what the user types into typed answers. The file decides what gets asked, which rules apply and when a tool runs. A tool that changes something cannot run until the user says yes to the exact inputs, and a file that breaks this rule refuses to load. It works with LangGraph. Use it in front of an agent you already run, or as the whole agent.

Good fit

Python teams building LangGraph agents that take real actions where a prompt-only policy is not good enough, as in banking, insurance or refunds. You get gates that work the same on any model, 2FA codes and amounts that only the user can supply, sign-off from a second person (for example compliance) before a step runs, and a plain-language procedure generated from the file that an auditor can read.

Consider something else when

Skip it if your agent only answers questions or runs read-only lookups, if you need open-ended planning, or if you don't work in Python. The step language has no loops, no parallel steps and no arithmetic, so complex logic means writing Python. Also skip it if you need multi-user authoring or conversations that survive a server restart out of the box.

Start here

Run pip install aigenos-flow, clone the repo, then run python examples/before_after/prove.py. No API key is needed. The same hostile input goes to a plain ReAct agent with the policy in its prompt, then to the same agent wrapped by Flow. The plain agent sends a $5000 transfer without a code and the wrapped one blocks it. Next, write one skill and wrap your agent with wrap_agent(agent, skills_dir="skills/", llm=llm, tools=[...]).

Cost and requirements

As of 2026-11-02 the library is free under Apache-2.0, and a paid team tier with SSO and hosting is planned. It needs Python 3.11+ and any LangChain chat model: Anthropic, OpenAI, Gemini or a local Ollama model. Expect one extra LLM call per message to decide which skill applies. The Studio editor runs from a clone of the repo and needs Node 20.19+ or 22.12+. React apps can use the separate @aigenos/flow npm package.

Limits to know

A skill only governs messages its classifier matches, so an agent you wrap that still holds a risky tool needs flow_only around that tool. A by: manager sign-off does not verify who answered, approval requests are kept in memory, and no one gets notified. Chat endpoints have no login or rate limit, and replies arrive whole rather than streamed. As of version 0.3.1, conversations are lost when the bundled server restarts, and the A2UI and MCP-UI adapters follow draft specs. The no-key proofs script the model, so they test the runtime, not any particular model.

Still unclear

• When will the planned paid team tier launch, what will it cost, and which features will need it?
• Does the experimental restart-safe storage path work with current LangGraph checkpoint releases?