For Linux · Weave Lite for macOS
Big work, broken into pieces
small models can actually finish.
Weave is an LLM chat app — and a suite of tools built around it. Chat with any model you like, local or remote. When the job is too big for one conversation, Weaves break it into isolated, bite-sized tasks a locally-hosted model can actually finish. Sandboxed tool use, MCP support, evaluations, skills, and more come built in.
One-time payment. Every device you own. All updates, forever. Currently in beta.
Chat you already know
A fast, no-nonsense chat interface for any model you point it at. Local or remote — your call.
Weaves, not walls of context
One big ask becomes many small, independent tasks. No shared context means no context bloat.
Yours, on your machines
No telemetry, no spyware, no bullshit. Runs locally, talks to the models you choose.
The core idea
What is a Weave?
A Weave is a collection of tasks. You, or the model, describe a large change, update, or project. Weave decomposes it so that each task is a single, isolated logical change: bite-sized enough that even a small self-hosted model can implement it correctly.
Describe the big thing
Type the large change in plain language. "Add dark mode," "port this module," "rework the billing flow."
Weave decomposes it
The work is split into discrete tasks, each scoped as one self-contained logical change.
Context stays isolated
Each task carries only what it needs. Nothing leaks between tasks, so the window stays tiny.
Small models do the work
Because every task is small and self-contained, a modest local model can finish it reliably.
Why isolation is the whole trick
Large models brute-force big tasks by holding enormous context. That's expensive and, increasingly, unsustainable. Weave flips it: keep each task so small and self-contained that the model barely needs context at all. The intelligence moves from the model's memory into the structure of the work.
Security Linux only
Sandboxed by default.
Letting a model run tools on your machine should never mean handing it the keys. On Linux, Weave confines every action to the project you're working on, and lets you keep a hand on the wheel for anything that writes.
Every tool call is sandboxed
Each tool call runs inside Bubblewrap, so it can only reach the files and directories of the project you're working on. Nothing outside the project is visible to the model — no home directory, no secrets, no surprises.
Every write is reviewable
Writes are captured through an OverlayFS layer instead of touching your files directly. You can optionally review every change by hand before it's committed to your actual project, or let trusted tasks flow through.
Bubblewrap and OverlayFS sandboxing are Linux features and are not available in Weave Lite for macOS.
Evaluations
Measure quality. Track it over time.
Run a prompt against one or more models, score every response with rubrics, and plot how quality changes between runs — so you always know whether a model swap or prompt tweak actually helped.
Multi-model comparison
Send the same prompt to multiple agents in one shot. Each response is independently scored by every rubric you attach.
Five rubric types
LLM Judge for nuanced quality, Contains / Does Not Contain for keyword checks, Regex for structured output, Tool Calls for agentic behavior.
Score trends
Tag your runs and Weave plots scores over time — so you can see whether a model change or prompt tweak actually moved the needle.
Reusable templates
Save a full evaluation setup — prompt, agents, rubrics, tags — as a template for consistent, repeatable runs after every model or prompt change.
Pricing
One price. No catch.
Weave Beta
Pay What You Want
one-time · in beta
- The full Weave application for Linux
- Weave Lite for macOS
- Run it on any number of devices
- All updates, forever
- No telemetry, no spyware, no bullshit
A theme for everyone
Features
Everything in the box.
One app, one price. Here's what you get the moment you install Weave.
- Familiar multi-model chat — OpenAI, Anthropic, or a local Ollama server, organized in folders with full history and token tracking.
- Weaves — decompose one big change into isolated, bite-sized tasks that run independently.
- Context isolation — every task carries only what it needs, so the window stays tiny.
- Sandboxed tools — on Linux, each tool call is confined to the project directory with Bubblewrap.
- Reviewable writes — on Linux, changes flow through OverlayFS so you can inspect them before they touch your files.
- Evaluations — run prompts against multiple models, score outputs with rubrics, and track quality trends over time.
- MCP support — connect MCP servers to give the model more tools.
- Automate & integrate — drive Weave from other apps over its HTTP API or built-in MCP server.
- Yours, on your machines — no telemetry, unlimited devices, all updates forever.
FAQ
Questions, answered.
Is Weave production-ready?
Weave is currently in beta. It's fully usable and we rely on it daily, but you may hit rough edges. Please keep backups and tell us what breaks at [email protected]. See the changelog for what's changed between releases.
What operating systems does Weave support?
Weave runs natively on Linux. macOS users get Weave Lite, which has the same chat and Weave workflow but without Linux-only sandboxing (Bubblewrap and OverlayFS).
Which models can I use?
Any model from OpenAI or Anthropic, or any local model served through Ollama. Point Weave at whichever provider you like — local or remote, your call.
Do you collect any of my data?
No telemetry, no spyware, no bullshit. Weave runs on your machine and only talks to the model providers you configure.
How do I get help?
Start with the docs, submit an issue in GitLab, or email [email protected].
Build for the world that's coming.
Bite-sized tasks. Isolated context. Small models that get the job done.
Get Weave — Pay What You Want