The Studio

The agentic canvas, in motion.

This is the product page that refuses to just describe itself. Watch a question get planned, retrieved, woven and cited - then read how each part works.

Browse example threads See pricing

Question

We're a 12-person team shipping an AI feature this quarter. Open-weight models or a proprietary API?

Decision memo

4 sources

Open-weight or proprietary: a build decision for small teams

For most small teams shipping this quarter, start on a proprietary API and keep an open-weight path open. The trade-offs that actually move the decision are below.

Time to first value

A hosted API puts a working model behind your feature in a day; standing up open weights adds infrastructure, autoscaling, and an on-call rotation before you serve a single request.1

For a team of twelve with a quarter-long deadline, that gap is the whole story: the fastest path to learning from real users wins, and self-hosting spends the first weeks on plumbing rather than the product.4

Cost at your scale

At low and spiky volume, metered API pricing is cheaper than the fixed cost of reserved GPUs plus the engineering time to keep them busy.2

The crossover arrives only once usage is high and steady enough to amortise that fixed cost. Most teams cross it later than they expect, and some never do.2

Control and lock-in

Open weights are the stronger answer when data residency, offline use, or a guaranteed exit path are hard requirements rather than preferences.3

The real lock-in is rarely the model; it is the prompts, evals, and glue built around one provider. Keep those portable and switching stays cheap.3

Recommendation

Start on a proprietary API and design for portability: keep prompts and evals model-agnostic, put the provider behind one interface, and revisit when steady volume crosses the cost line.2

woven in 1.84s · 4 sources · 412 tokens

Anatomy of a weave

Four moving parts. Nothing happens off-screen.

warp

The plan

Before any text, the agent lays a structure - the sections the answer will need. Set first, so the result is organised, not a wall of prose.

weft

The sources

Retrieved threads enter from the edges and interlace with the plan. Each weft is one source, ranked and held in view rather than hidden.

crossing

The resolve

Where a source crosses a section, content resolves in place - a sentence streams in, grounded in exactly the thread it sits on.

shuttle

The provenance

The accent shuttle marks the agent's focus, leaving a citation behind it. Every claim ends up linked to the source it came from.

Built to be checked

The point is not a faster answer, but one you can audit.

A finished artifact: Not a chat log. A structured brief with headings, a recommendation, and a clear shape you can hand to someone else.
An open trace: The plan, the retrieval, the order of work. The agent shows what it did, step by step, in the machine voice.
Named sources: Every weft is a real entry in the source list, and every claim links to it. Provenance is the default, not a setting.

Read a few weaves people have already run.

Browse threads Why we built it