Access is explicit.
Files, tools, browser control, and cloud context sit behind visible controls. The agent does not quietly expand its own reach.
Local-first AI workbench
Fae
Fae helps you think, research, write, code, inspect, and act inside real projects while keeping context, tools, approvals, diffs, and recovery paths visible.
Product
A local workbench for AI that can act.
Fae connects conversation to project context, tool access, process artifacts, file diffs, memory, and model choice. State stays close to the work it affects.
Process artifact, completion state, file diff, and local work context.
Reviewable work
Changes stay close to their evidence.
When Fae completes a task, the result is paired with the process and files it touched. Review becomes part of the workflow, not a separate cleanup step.
Control model
The trust model is part of the product.
Fae treats agent power as something you grant, inspect, and can roll back. The app shows what the agent can touch, what it changed, how the result was checked, and what can be undone.
Work leaves artifacts.
Process logs, changed files, diffs, and summaries stay available after the turn, so completion is backed by something you can inspect.
Recovery comes first.
Writing turns capture a recovery path where possible. When Fae cannot recover something, the product should say so plainly.
Working interface
The controls stay where the work happens.
Slash commands, project mode, local file access, model selection, and process state are part of the same surface. You do not have to leave the workbench to understand what mode you are in.
Command surface over an agent run.
System surfaces
Settings and history are product surfaces.
Behavior, subagents, voice, models, integrations, files, memory, and data controls.
Local activity, agent runs, changed files, model usage, and workspace history.
Early access
For people who want AI in their actual work.
Fae is pre-release. Early access is for people willing to test it on real projects and tell us where the trust model helps, where it slows them down, and what still feels unclear.
Windows desktop first.
Best fit: repositories, research folders, browser workflows, and privacy-sensitive work.
Local models through Ollama. Cloud providers are optional and context-gated.