SinWeave
v1.0.0-α
Download
In August 2026 DeepSeek open-sourced Harness — the agent runtime their own models were built to run in. Not a model, not an API: the loop that reads your files, calls tools, runs commands, and keeps going. It is MIT-licensed, and it is coming to SinWeave.
vitest checkout ten times to reproduce.
3 of 10 failed on cart.total — a shared fixture is
mutated between cases. Patching the fixture, then re-running.
18- const cart = baseCart 18+ const cart = structuredClone(baseCart) 31 expect(cart.total).toBe(4200)
Most agent tools hard-wire the loop and let you swap the model. Harness inverts that: every layer is a plugin, including the loop itself. That is the whole reason it is worth adopting rather than reinventing.
DeepSeek V3.1 is the default, but inference is a plugin like everything else. Point it at Anthropic, OpenAI, Gemini, OpenRouter, or the Ollama instance already running on your laptop.
Commands run under an enforced file policy — read-only, workspace-write, or full access if you insist. Landlock and bubblewrap on Linux, Seatbelt on macOS, restricted tokens on Windows.
Context compaction, spill-to-disk, session storage, and checkpoints are first-class packages. A run that takes forty minutes and two hundred tool calls does not fall over halfway.
Fan work out to child agents, each with its own budget and toolset. Plans, goals, and todo state are tracked by the runtime rather than improvised inside a prompt.
Model Context Protocol servers attach directly. Language-server diagnostics feed the agent real type errors. Hooks let you intercept any tool call before it runs.
Every step is recorded: which plugin fired, which tool it called, what came back. When an agent does something surprising you can read exactly why instead of guessing.
SinWeave already does chat, inline edits, and multi-file diffs well. Harness is about the jobs that are too long or too branching for a chat window.
The plan is deliberately boring: the runtime sits beside the editor, speaks a documented protocol, and never becomes something you cannot turn off.
No separate install, no daemon to babysit, no account. It starts with the editor and stops with it, and you can disable it entirely in Settings.
Whatever you have already configured becomes the harness's model layer. Nothing new to paste, and no provider you did not choose yourself.
Every step appears as it happens and every write still lands as a reviewable diff. The sandbox policy is yours to set — read-only, workspace-only, or open.
▸ why are the checkout tests flaky? → sandbox: workspace-write · model: your key → run vitest checkout --repeat 10 → read src/test/fixtures.ts · 3 callers → edit fixtures.ts · clone the base cart → run vitest checkout --repeat 10 ✓ 10/10 green · 14 steps · 2 files · review the diff ▸
Early access
We will send one email when the Harness build is ready to try, and a second if we need testers before that. No newsletter, no drip sequence, no sharing your address with anyone.
No. Harness is MIT-licensed software that happens to have been written by DeepSeek, in the same way SinWeave is a fork of software written by Microsoft. The model layer is a plugin — DeepSeek V3.1 is the default because that is what upstream ships, and you can replace it in one setting.
Only if you choose a DeepSeek model, and then it goes directly from your machine to their API with your own key. Pick Anthropic, OpenAI, or a local Ollama model and DeepSeek never sees anything. The runtime itself is code running on your laptop, not a service.
Because the hard parts — sandboxing across three operating systems, context compaction, session recovery, subagent budgets — are solved there and battle-tested by a very large number of users. Writing a worse version of that ourselves would cost months and buy nothing.
No date yet. Harness is still a developer preview and its own authors warn about breaking changes, so we are tracking it rather than racing it. The waitlist is the honest answer: you will hear the day there is a build worth using.
No. SinWeave stays free and Harness is MIT-licensed. You pay your model provider directly, or nothing at all if you run local models. Long agentic runs do burn more tokens than a chat message, so keep an eye on your provider dashboard.
Yes — it is open source and runs standalone today. That is a separate thing from the SinWeave integration, which is what this page and the waitlist are about.