><> sprat

TERMINAL RESEARCH AGENT

Research the problem.
Shape the product.
Show the evidence.

Give Sprat an idea or a market question. A small team of research roles investigates it with real, cited sources, challenges the assumptions, and hands back a product brief and a validation plan you can act on — showing what every role did and what it cost.

Get early access See where it is

In development, Phase 1. Not installable yet — the build is happening in the open.

WHAT YOU ASK IT

> I have an idea.

Help me decide who it is for, and what to build first.

> Research this market.

Find a worthwhile opportunity, and the evidence against it.

> Here is my repository.

Given what already exists, what should I build next?

IN THE TERMINAL

> I want to build software for independent gyms in the UK

⏺ PM  Before researching, three things change the direction:
      1. Owner-operated gyms, or small chains (2–5 sites)?
      2. Is the goal revenue, or a tool for your own gym?
      3. Any constraint on building mobile apps?

> owner operated, revenue, web first is fine

⏺ Stage: Research · standard depth · 2 researchers

⏺ Researcher·1  web_search("independent gym management software UK")
  ⎿  10 results · 4 new sources recorded

⏺ Researcher·2  web_fetch(reddit.com/r/gymowners/…)
  ⎿  Fetched 3.1 KB · 2 claims recorded

⏺ Researcher·1  Finished competitor matrix (7 products, 12 sources)

✢ Researching… (2m 14s · ↓ 48.2k tokens · esc to interrupt)
  ⎿  ☒ Intake
     ◼ Research          Researcher·1 done · Researcher·2 fetching
     ☐ Analysis
     ☐ Directions
     ☐ Brief + validation plan

⏵ normal mode · $0.84 · 31 sources

Inline, in your scrollback. Not a full-screen dashboard.

WHY TRUST THE OUTPUT

Every claim carries its source

URL, publisher, publication date, retrieval date. Click through and check the work.

Counter-evidence is recorded too

Claims store contradicting sources alongside supporting ones, with a confidence level and a reason.

Unknowns stay unknown

A missing figure scores unknown, never zero and never an average. Evidence coverage is shown beside every score.

Gates are enforced by code

A role cannot skip a stage or write a later artifact because the model felt confident. Application code checks, not instructions.

Nothing is invented

Sprat never fabricates interviews, customer quotes, demand figures or willingness-to-pay data, and never claims product-market fit.

You see what it cost

Per role, per call, including failed and cancelled ones. Money is capped before the call is made, not apologised for afterwards.

Your research never leaves your machine.

Sprat runs locally, on your own Anthropic key. Missions, sources, claims and transcripts live in a SQLite database on your disk. There is no cloud control plane, no sync, and no telemetry — not opt-out, not opt-in, not present.

An account here holds your entitlements, not your work.

WHERE IT IS

Phase 0

Feasibility

Prove the central bet and the risky parts before building the product.

The build record is public, including the parts that are not working and the one experiment that could still change the shape of the product.

Read the milestones

Get early access

One email when there is a binary to install. No newsletter, no drip sequence.

Join the list