After 262 commits on a fork of OpenClaw, we stopped calling it a fork. It’s a different animal now.
TinkerClaw started as a personal itch — we wanted to see where our AI tokens were going. Six weeks later, it has its own cognitive architecture, a real-time dashboard, and a memory system that vanilla OpenClaw doesn’t offer.
Why We Forked
OpenClaw is excellent at what it does — running AI agents across multiple channels. But when you run agents 24/7 in production, you hit walls that a general-purpose framework doesn’t solve:
- No cost visibility. A single deep Opus session can cost $20+. You don’t find out until the invoice. We needed real-time tracking.
- No persistent memory. Every session starts blank. For an agent that manages your crons, files, and messages — that’s unacceptable.
- No self-improvement loop. An agent that runs the same mediocre cron job on Day 30 as Day 1 is wasting your money.
What TinkerClaw Adds
Tinker UI — Real-Time Token Dashboard
A Vite + Lit web dashboard that shows you exactly what’s happening inside your agent. Context window treemaps, response cost breakdowns, per-provider tracking — all live, not after the fact.
ENGRAM + HIPPOCAMPUS — Cognitive Architecture
A memory system inspired by how biological memory works. ENGRAM handles episodic memory (what happened), HIPPOCAMPUS handles consolidation (what matters). Your agent wakes up remembering yesterday, last week, and that one config crash three weeks ago that it should never repeat.
Self-Improving Crons
Cron jobs that review their own output and rewrite their prompts for next time. We call it the META pattern. Day 1 mediocre → Day 30 expert. No human intervention needed.
Budget-Aware Model Routing
Automated work runs on flat-rate models. Metered models only fire for cross-model review or explicit user requests. Budget pressure adjusts automatically: >85% caps at Sonnet, >95% drops to Haiku. No surprises on the bill.
The Numbers
- 262+ fork commits ahead of upstream
- 8 ClawHub skills published and open source
- 32 operational lessons documented in the Field Guide
- Context window from 23.5KB → 12KB
- Zero VC money. One developer. Problems worth solving.
Get Started
git clone https://github.com/globalcaos/tinkerclaw.git
cd tinkerclaw
pnpm install
pnpm buildFull setup in the README. The Field Guide covers everything from first boot to production.
Read More
- What Running an AI Agent 24/7 Actually Costs
- We Taught Our AI to Teach Another AI
- ⭐ Star TinkerClaw on GitHub
Clone it. Break it. Improve it. That’s the point.