TinkerClaw started in February 2026 as a small itch: I wanted to see where my AI tokens were going. Eight months later it is the system I work in every day. It has more than 3,600 commits that upstream OpenClaw does not have, and it no longer feels like a patch set. This page is the tour I would give a friend sitting next to me: what each panel shows, why it exists, and what it still gets wrong.
Why fork at all
OpenClaw is good at what it was built for: one agent, many chat channels. Run it around the clock for real work and three walls show up fast.
- You cannot see the cost. A long session on a frontier model can spend more than lunch, and you learn about it on the invoice.
- The agent forgets. Every session starts blank, which is fine for a chatbot and bad for something that runs your scheduled jobs, files and messages.
- Nothing gets better on its own. The same mediocre job runs on day 30 exactly as it ran on day 1.
Every panel below exists because of one of those three walls.
The Tinker UI at a glance
The Tinker UI is a web page, built with Vite and Lit, that talks to the gateway. The chat sits in the middle. Each tab along the top is its own conversation with its own memory and its own bill. The left rail opens full-page views such as the Recipe Book and the cron board. The right rail is where the panels live, and they always describe the tab you are looking at.
Models: pick the brain and how hard it thinks
Two questions decide the cost of every answer: which model, and how much it thinks. The Models panel puts both under your thumb. Each coloured tile is a model, coloured by provider. About 80 are connected, from Anthropic, OpenAI, Google, xAI and several Chinese labs.
Pin a tile and you set its effort yourself, from Minimal to Max. Leave it on Auto and the effort row turns into a bias dial, from budget to smart, and the router decides each turn. The longer story is in the effort × model slider post.
Thalamus: who answered, and why
On Auto, a router I call Thalamus reads each turn and picks a model and an effort for it. Its card shows what is running right now (here Opus 5.5, on Auto, writing the main reply) and how big the job is. Its rule is simple: a fix that keeps failing gets a stronger model, not more thinking from the same one. The design is in the Thalamus paper.
Context window: what the model is actually reading
Every call sends the model a pile of text: the system prompt, files, skills, tool results and the conversation itself. The coloured bar shows that pile, split by kind, against the model’s window. In this shot it is 45,000 tokens out of a million.
Below it are counters for this call and for the whole session: turns, compactions, evictions, tokens dropped. The Evict and Compact buttons let you shrink the pile by hand. The call timeline at the bottom draws each model call of the current prompt as it happens.
When the window does get squeezed, the memory layer keeps pointers to everything it moves out, so a detail can be fetched back instead of lost. That is the idea behind Total Recall; a concept index, described in Instant Recall, makes the lookup fast.
EEG: where the effort and the money went
The EEG draws every model call as a stroke. Left to right is the effort it ran at. Colour is the provider. Height is what it cost, on euro gridlines that rescale to the tab. Yellow lines mark each prompt I sent, and helper agents branch off the main trace and join back when they finish.
It exists for one reason. My agent chooses its own model and effort, and I wanted to check that cheap jobs really go to cheap models. The full explanation, including what it still undercounts, is in Watching an Agent Think.
Recipes: the agent’s playbooks
A recipe is a markdown playbook: numbered steps, constraints, safety notes, and the failures it has already survived. When a task matches one, the agent follows it instead of improvising. The Recipe Book view lists them all, searchable, with an editor. The Recipes panel in the right rail shows the plan of the task in progress: which step it is on and which helpers are working. More in The Recipe Book.
Amygdala: a reflex before the agent acts
The Amygdala is a small, fast layer that looks at what the agent is about to do at four points: when my prompt arrives, before a tool runs, after it, and at the end of the turn. Twenty hard rules, written as code, stop the obvious mistakes: deleting a file I did not name, wiping a disk, dropping a database, force-pushing over the main branch, or reading a password nobody asked for.
A cheap judge model handles the grey cases. Most of it runs in shadow mode: it records what it would have done, and a rule only starts acting after enough reviews show it helps more than it gets in the way.
I kept this screenshot honest. The red line says the judge had been silent for over three hours that morning, with the hard rules still on. A safety layer that hides its own outage is worse than none. The paper is AMYGDALA.
Exec mode: the cockpit
A switch in the top bar slides in the Exec panel on the left. It has three tabs. Pulse shows the numbers I watch: GitHub activity on TinkerClaw, ClawHub views and installs of the skills, open issues and pull requests.
Today holds a seven-day calendar strip and the task board. Tasks show a yellow dot while the agent is working on them. Crons lists the reports the scheduled jobs left overnight. Today and Crons are full of my own life, so they stay out of this page.
Crons: work that happens while I sleep
Twenty-two scheduled jobs live on the cron board, eighteen of them switched on, and the history held 2,711 runs on the day I took this. A morning briefing, a security check, a memory cleanup, a scan of other forks for ideas worth borrowing, a nightly pass that turns the day’s mistakes into fixes. Each run shows which model did it and how many tokens it took.
The jobs also learn. Each one keeps a lessons file that the next run reads and adds to. Changes to the instructions themselves go through me first: I learned the hard way that a job allowed to rewrite itself unchecked drifts. More in self-improving cron jobs.
Smartness × Cost: which model is worth it
The AI Analysis button in the Models panel opens this map. Up is smarter, using the Artificial Analysis intelligence index. Right is more expensive per million output tokens, on a log scale. Each line is one model across its effort levels. It is what Thalamus works from, and I publish it as a page anyone can use: AI Analysis, the Smartness × Cost map.
Under the hood
A few parts have no panel of their own but carry the rest.
- Claude Code bridge. Chats can run through the Claude Code command line, on the subscription I already pay for, instead of per-token API billing.
- Moral code. A short set of rules (truth before agreement, privacy, ask before anything you cannot undo) is handed to every session at its start and after every compaction. It outranks later instructions.
- Fractal reflection. After a turn, a background turn asks what it taught and writes the lesson down. So far it has not paid off; the post-mortem says why.
- Memory. Daily notes, a curated long-term file and a searchable bank that every session shares. It is how the agent knows on Monday what went wrong on Friday.
The numbers, October 2026
- More than 3,600 commits that upstream OpenClaw does not have
- 26 skills published on ClawHub
- 107 recipes in the Recipe Book
- 22 scheduled jobs, 2,711 runs recorded
- About 80 models to choose from, across a dozen providers
- One developer, no outside money
What it does not do yet
- The EEG and the cost figures count output tokens only, so every euro amount is an underestimate.
- Costs for most providers outside Anthropic are estimates.
- The Amygdala judge can go quiet, as the screenshot shows. The hard rules keep working when it does, and the panel says so, but the judge should not stall.
- Setup still assumes you are comfortable in a terminal.
Try it
git clone https://github.com/globalcaos/tinkerclaw.git
cd tinkerclaw
pnpm install
pnpm buildFull setup is in the README. The Field Guide covers the lessons from the first months, from first boot to running it around the clock.
Everything on the site, by topic or date: Browse the Tinker Zone