mrdiamondballz/blackfin-agent

Blackfin is a Python tool for running AI agents on your own computer instead of a cloud service. It is built to feel familiar to people who already use OpenClaw, but it adds stricter rules about what an agent is allowed to do and keeps a record of every step an agent takes. With Blackfin you set up named agents that each get their own private workspace, memory, and rulebook, called a policy. A policy controls things like whether an agent can read files, run shell commands, or make web requests, and by default agents start out limited rather than fully open. You can connect agents to Telegram or Discord so they can chat with you there, or run them from the command line. Agents can also use different AI model providers, including Ollama, OpenAI compatible services, OpenRouter, or vLLM. A notable part of this version is a system for trusting or rejecting add-on skills before they run. Skills are small packaged capabilities that can be installed into an agent, and Blackfin checks who created a skill and whether that creator is trusted before allowing it to run, rather than trusting everything automatically. There is also a simple text based dashboard you can open in a terminal to see recent activity, memory entries, and scheduled jobs at a glance. Every action an agent takes is logged as a trace, and memory is searchable, which is meant to help a person understand what an agent actually did rather than guessing. Blackfin only depends on the Python standard library, so it does not require installing many extra packages to get started. The project includes documentation covering its security approach, its overall architecture, and how to move over from OpenClaw. It ships with 222 automated tests. The license is described as source-visible, meaning the code can be viewed, but no open-source license is granted by default, so check the license terms before reusing it commercially.