The LaForge Roadmap
LaForge is software for running AI and human teams through auditable workflows — every decision, execution, and test result logged and verifiable. The claim on the tin is simple: prove what your AI agents did.
We build it in public, with itself, and this page is the order of what comes next. No dates — sequence is a promise we can keep, and dates are a promise software breaks. Milestones ship when their evidence exists, and each one gets written up honestly when it lands.
Shipped — install anywhere#
One command on the homepage, no root, no account. A signed release, a verified channel, and an installer that refuses any byte it cannot prove is the byte we published — demonstrated by sabotaging it seven different ways and watching it decline every one. Uninstall removes what was added and touches nothing of yours, and it checks before it believes itself. The story of that day is on the blog.
Next — the record gets verbs#
LaForge keeps a product record: every feature, decision, and piece of evidence in a durable, readable home. Next it becomes fully operable — creating work, recording decisions, and briefing an agent for a task become single commands, and the console reads live from the record. The proof we're holding ourselves to: LaForge's own backlog, managed entirely by LaForge. Decisions that need a human get raised to the operator with their context written out and their options laid flat — including the option to write in your own.
Then — a record you can trust with everything#
Before a record holds the only copy of anything, it earns that trust: tested backup and recovery, migration that survives its own growth, search across everything ever decided, and instrumentation that watches the record's own health — so the system can detect the conditions that make systems like it necessary. Defects and acceptance criteria become first-class records, so "which requirements are proven?" becomes a query instead of a meeting.
The proof — no one certifies their own work#
This is the reason LaForge exists, and we will not promote it as a product until this ships. Separation of duties, enforced by structure rather than policy: the agent that builds a thing cannot be the one that certifies it, every verification check must demonstrate it can fail before its passes count, and the depth of evidence required scales with the risk of the change. When this lands, "prove what your AI agents did" stops being a pitch and becomes a demonstration.
The harvest — migration, and the evidence pack#
Our own production history — years of items, rulings, and evidence from the systems that taught us this method — migrates into LaForge, and those systems retire. Then the artifact all of this exists to produce: the evidence pack. A scoped, point-in-time, self-contained record a compliance team can hold — what was claimed, what was proven, by whom, and what remains honestly open.
Stations — your own hardware, one command away#
A remote machine — a workstation, a DGX on your bench — enrolls as a model host through a daemon that only ever connects outward; nothing on your network starts listening. Image, audio, and video generation become installable providers under one capability model: a hosted service like OpenArt and your own GPU are the same shape to the caller. Starting a model on a machine down the hall stops being an ssh session and becomes a command.
Beyond#
Remote operation that never opens an inbound port. Integrations that push decisions and risks into the systems your company already uses — your tooling stays your tooling, and the agents never need to know it exists. A plugin ecosystem with signed, verifiable distribution. And the authority model for the second human: the reviewer, the auditor, the operator who may rule but not certify.
Follow along on the blog, or start the conversation about what accountable AI engineering looks like in your organization at VRAISON. Install is one line, on the homepage.