← All work
AI agents · Personal tool · Claude Agent SDK
AgenticBrain
My own control center for Claude agents. I give it a task from my desk or my phone; it picks a model, works in a sandbox, has an independent reviewer check the result, and remembers what it learned. It also proposes improvements to itself, and nothing changes until I approve it.
- Role
- Solo: design and build
- Stack
- Python · FastAPI · Claude Agent SDK
- Runs
- Docker, phone access over Tailscale
- Safety
- Budget cap · kill switch · approvals
What it does
- Model routing: each task goes to Haiku, Sonnet or Opus depending on how hard it looks, so simple jobs stay cheap.
- Sandboxed tools: the agent reads, writes and runs code only inside its workspace folder, plus web search and fetch.
- Independent self-check: a separate reviewer checks the finished work and sends it back for a fix round when something's wrong.
- Memory: a local vector store of durable notes; the relevant ones are pulled into each new run automatically.
- Brain Warden: keeps that memory clean by proposing merges, prunes and new standing rules. Uncertain changes wait for my approval, and applied ones can be rolled back.
- Self-improvement loop: the agent proposes upgrades to its own code, and I approve, run or dismiss each one.
- Guardrails: a spend meter with a hard monthly cap, and a kill switch that stops live and scheduled runs.
- Tools: MCP connectors can be switched on per project or per agent type, so each chat loads only what it needs.
- Scheduling: recurring jobs like a nightly dependency audit, with stall detection.
Screenshots show the real interface running on invented demo data. It's a personal tool, so my own projects and spending aren't shown.
Inside
It proposes, I approve
Let's build something.
I'm open to full-time and contract roles in AI/agent engineering, mobile, and full-stack TypeScript, remote or hybrid around NYC and Philadelphia.