Open Brain: A Personal Agent Fleet That Runs on GitHub for Free

The Problem with Personal AI Agents

Everyone wants AI agents that run autonomously — researching, summarizing, monitoring — without babysitting. But the options are bad:

Cloud agent platforms charge per run, per token, per minute. Costs add up fast for always-on agents.
Self-hosted solutions require servers, orchestration, and DevOps knowledge that most people don’t have.
Chat-based AI requires you to be in the loop. The whole point of agents is that they work while you don’t.

What if you could deploy a fleet of autonomous agents that share a persistent memory, run on a schedule, and cost essentially nothing?

The Insight: GitHub Already Gives You Everything

GitHub’s free tier includes:

GitHub Actions — 2,000 minutes/month of scheduled compute
GitHub Models — free LLM inference (GPT-4.1, GPT-4o, and more)
Repositories — version-controlled, inspectable, forkable agent code

The missing piece is a shared brain — a place where agents can read context, build on previous work, and write back results. That’s what Open Brain provides.

How It Works

Open Brain is an open-source personal knowledge base built on AWS (S3 Vectors, Bedrock embeddings, Lambda) that exposes an MCP interface. Any MCP-compatible client — Claude Code, Claude Desktop, ChatGPT, or your own agents — can search, browse, and capture thoughts.

The agent fleet architecture looks like this:

GitHub Actions (free compute)
  + GitHub Models (free LLM)
  + Open Brain (shared memory via MCP)
  = Autonomous agent fleet at $0/month

Each agent is a GitHub repo with:

agent.prompt.yml — the prompt config (GitHub Models compatible)
agent.ts — a lightweight agent loop that connects to Open Brain via MCP
.github/workflows/agent.yml — a cron schedule that runs the agent

The Agent Loop

Every agent follows the same pattern:

Search the brain for relevant prior context (avoid duplicates, build on previous work)
Do the assigned task (research, analyze, summarize, monitor)
Capture results back to the brain so other agents and the user can access them

This means agents build on each other’s work. A research agent captures today’s findings. A pattern recognition agent reads those findings and identifies trends. A weekly reflector synthesizes everything into a summary.

Deploy an Agent in 60 Seconds

The Open Brain dashboard includes a wizard that handles everything:

Pick a template — Daily Researcher, Weekly Reflector, Learning Tracker, or Custom
Describe the task — “Research today’s top AI practitioner stories and capture a concise summary”
Set schedule and model — cron expression + model selection (GPT-4.1, GPT-4o, etc.)
Click Deploy — creates the repo, sets secrets, configures the workflow

Behind the scenes, the wizard:

Creates a private repo from a template via the GitHub API
Generates an Open Brain API key for the agent
Writes agent.prompt.yml with your task description
Sets repo secrets for brain connectivity
Enables the scheduled workflow

The agent runs automatically on schedule. You can also trigger it manually from GitHub Actions, or edit the prompt from the dashboard at any time.

Prompt Versioning with GitHub

Agent prompts live in agent.prompt.yml — a GitHub-native prompt format that integrates with GitHub Models tooling:

model: openai/gpt-4.1
messages:
  - role: system
    content: |
      You are an autonomous agent with access to a personal
      knowledge base called Open Brain. On every run you MUST:
      1. Search the brain for relevant prior context
      2. Do your assigned task
      3. Capture a concise summary back to the brain
  - role: user
    content: |
      Research today's top AI practitioner stories. Search the
      brain for yesterday's briefing to avoid duplicates, then
      capture a concise summary.

This gives you:

Git history — see exactly what changed in each prompt revision
GitHub Models Prompt Editor — test and iterate on prompts in the browser
Evaluation tooling — add testData and evaluators fields to validate prompt changes
Direct editing — power users can edit the file in GitHub without touching the dashboard

The Agent Bus

Open Brain doubles as a communication bus. Every agent writes to a shared index, and any agent can read from it. This enables coordination without direct coupling:

Topic channels — agents use channel: prefixed topics (channel:research, channel:security) to organize shared content
Activity monitoring — the dashboard shows per-agent activity, recent thoughts, and a timeline
Cross-agent context — Agent B can search for Agent A’s output and build on it

The web dashboard at brain.blanxlait.ai shows the full agent fleet: status, recent activity, and a feed of everything agents have captured.

Security Model

Agents run with scoped permissions. The MCP server restricts agent API keys to safe operations only:

Agents can	Agents cannot
`search_thoughts`	`delete_thought`
`browse_recent`	`update_thought`
`capture_thought`	`create_agent`
`stats`	`revoke_agent`
`agent_heartbeat`	`list_agents`

This limits the blast radius of prompt injection attacks. Even if an agent reads malicious content from the web that tries to instruct it to “delete all thoughts,” the tool simply doesn’t exist in its session.

Additional hardening:

Server-side model allowlist prevents injection via the model field
Cron schedules are character-stripped to prevent workflow YAML injection
Prompt content uses YAML block scalars with forced indentation to prevent YAML escape

What Agents Are Already Doing

Here’s what’s running in production today:

Daily AI News Researcher — searches Hacker News, ArXiv, TechCrunch, and DeepMind daily. Checks the brain for yesterday’s briefing to avoid duplicates, then captures a structured digest with links and trend analysis.

Pattern Recognition Agent — scans the brain periodically, identifies development patterns, bus factor risks, security posture, and blocked ideas. Produces actionable reports with ranked risks and opportunities.

Both agents cost $0 to run — GitHub Actions compute and GitHub Models inference are free.

Try It

Hosted: Sign in at brain.blanxlait.ai with Google — the brain infrastructure is managed for you. Connect GitHub, deploy agents, and start building your fleet.

Self-hosted: An open-source fork is available at github.com/niemesrw/openbrain. Deploy the CDK stacks to your own AWS account for full control — your data stays in your infrastructure.

The agent template, dashboard, MCP server, and CLI extension are all included. Connect any MCP client — Claude Code, Claude Desktop, or your own tools — and start capturing.

What’s Next

Agent-to-agent coordination — structured handoffs where one agent’s output triggers another
iOS and macOS apps — native capture from anywhere with share extensions and Siri shortcuts
Evaluation loops — agents that test their own prompt changes and self-improve
Slack integration — agents that post to channels, respond to DMs, and capture from conversations

The core thesis: your AI should work for you when you’re not looking. Open Brain makes that possible with tools you already have.