AI work nodes · open source · verifiable

Make AI actually finish the work,
and prove that it did.

Every done needs a receipt: results, process, and what was missed can be checked. Start with the beta checkup to compare your own answers with a scan, install a local task engine, or try the LINE lab first.

Install a local task engine →Tried it or want to join? Check in at Future Village →

Help each person grow their own human–AI partnership

Here, the partnership means the Captain plus the primary AI: the person decides, and the AI helps execute and verify

Find the gaps first, then add tools, learning materials, or hands-on support. A prepackaged brain is not the offer

AI checkup: public beta

Give the repo to your AI and read the instructions first. Compare your self-assessment with the scan of your computer

It can misjudge. A person must review the result. Remove private details before sharing; do not post raw logs

Read the checkup instructions and known misjudgments →
MINI BRAIN

See how it works, then decide whether to install it

Send it a task, let it run in the background, get the result back, and watch progress in the browser. Task state stays in files on your machine. You can open them. Open source, MIT. Default mock mode does not need an API key. To make Claude CLI actually run, you use your own Claude account.

Path A · I want to run it myself

DIY: clone it and run it

If you can open a terminal and copy-paste, you can start. Default mock mode does not need an API key. To make Claude CLI actually run, you use your own Claude account. One prompt can ask your AI coding agent to install it.

Read the one-prompt install →
Path B · I want to see it in action first

Tour the AIWFF R&D workspace (Machine A)

See how a task is assigned, leaves a result, and is checked afterward. This is the Captain's own R&D workspace. The public local task engine is a starting point, not a packaged copy of Machine A.

Watch the full tour on the ZAX Taiwan channel →
FEATURED OPEN SOURCE

Featured open source: build the engine first, then add the guardrails

These are tools and entry points published after iteration in the AIWFF R&D workspace (Machine A). Start with aiwff-runtime for the engine, soplint for rule checks, or line-persona for a LINE twin. See the cards for stars and update dates; the copy does not freeze those numbers. It is not me saying the project is alive. GitHub says it.

aiwff-runtime

★ 6 · MIT
Last updated

An agent runtime where every "done" leaves a receipt: install it on your own machine. Default mock mode does not need an API key. To make Claude CLI actually run, you use your own Claude account. Then decide whether to connect a real worker.

Method: A local file bus records tasks, progress, and outputs, so users can watch status in the browser and still inspect the file evidence later.

Result: Packaged into the local task engine page, install guide, and public repo so it can be adopted from zero.

Proof: Tests are on GitHub Actions. The link is here. Pass or fail is whatever that page shows now.

View entry →

soplint

★ 46 · MIT
Last updated

Static SOP audit for AI work nodes. It catches instruction drift.

Method: Turns SOPs into scannable static rules, audits node outputs line by line, and catches drift.

Result: Open source, with automated tests on every push. The star count on this card is live.

Proof: Tests are on GitHub Actions. The link is here. Pass or fail is whatever that page shows now.

View repo →

line-persona

★ 15 · MIT
Last updated

Fill in three files and you get an AI twin that lives on LINE, talks like you, and runs on the model you choose.

Method: Keys and model go in .env, your voice in profile.md, your data in knowledge.md. Switch between cloud and local models freely. If you don't code, have your AI read AGENTS.md and set it up for you.

Result: v0.2.0 is open source. The lab account is live on LINE.

Proof: Chat with a live one on LINE

View repo →

dig-loop

★ 0 · MIT
Last updated

An AI digs one domain each day for problems that are not solved yet. Solved items are not collected again. Finished digs go on the solved list.

Method: Each find is checked against the solved list. New ones come back. A person decides what is worth doing.

Result: A dig counts only after it is written down. The next round does not collect the same hole.

Proof: Header check script. Clone it and run it yourself.

View repo →

grok-bot-routines-tw

★ 0 · MIT
Last updated

Plenty of guides teach you to schedule a bot. This one adds the check that it actually finished.

Method: Hand the repo to your AI, or to Grok Bot. Both have to match a row on the automation list.

Result: The AI saying it is built does not count. The row has to be on the list.

Proof: The AGENTS.md to read before handing this to an AI

View repo →

execution-proofs

★ 6 · MIT
Last updated

MCP telemetry gateway. Forces agents to prove done with real files and timestamps.

Method: Built as a local MCP telemetry gateway that intercepts "done" claims and requires real files plus timestamps.

Result: Open source, test suite included. Clone it and run it yourself.

Proof: Test file: clone it and run npm test

View repo →
See all open source projects →
LINE LAB

Do not want to install yet? Talk to a running LINE twin first

Talk to a running LINE twin. The lab account is live on LINE.

Add on LINE: @395jcpsb