PromptPotter
Try it — it’s free
Connector Your infrastructure

your pipelines, your business logic, your LLM calls

One call: your input, an LLM, your answer.

LLM

Three steps: your input, a web search, an LLM, a ranking, your answer.

retriever web search
LLM
tool ranking

What a connector is

  • Everything it could try: Every version it writes is a point here, coloured by how well it scored.
  • One round: Write versions, score them, keep what worked. The best one starts the next round.
  • When it stalls: If a round stops improving, it steps back and tries a different angle.
Measured, not guessed Live

Show it what a right answer looks like. It works out how to get there.

Try it — it’s free
Free & open source

The whole thing is on GitHub.

No license keys, no seats, no paywall — the full source is public. Clone it, read it, run it.

View on GitHub Read the docs
The routing wave

Everyone is choosing a cheaper model. We fix what gets sent to it.

Model routers are having a moment, and they earn it: send each call to the cheapest model that can still do the job, or to a free one, and the bill falls. Not one of them touches the thing being sent.

Model routing

  • Switchyard Sends each step of an agent to a cheaper or a stronger model.
  • OpenRouter One endpoint in front of hundreds of models, priced per call.
  • LiteLLM One OpenAI-shaped API in front of a hundred providers.
  • Portkey A gateway with fallbacks, retries and 1,600+ models behind it.
  • Devin Fusion Pairs a frontier coding agent with a cheaper sidekick mid-task.
  • Kong The same routing, applied at the API gateway you already run.
  • Bifrost A fast self-hosted gateway with adaptive load balancing.
  • RouteLLM Easy questions to the small model, hard ones to the big one.
  • Not Diamond Picks the model most likely to get this particular call right.
plus the pipeline

Model routing + pipeline optimization

PromptPotter It changes all of it, and measures each change:
  • the wording of the request
  • the examples you show it
  • thresholds and retries
  • which model runs each step
  • any pipeline setting you expose

Routers pick who answers. This fixes what they are asked, and proves the difference with a score on your own examples. The two stack rather than compete: OpenRouter is the default provider here, not a rival.

The operator console

Watch every round, right in your browser.

While it runs, a console streams what it is working on, the version being scored right now and how every round came out — no extra tooling, no setup.

Take the full tour →
localhost:8001
production · 84% acc
Input
Query
TermNorm
web search AI agent ? your node
Output
Answer
customer-tickets
round 4 running
Round 3 done — 0.86, up from 0.79. Step-by-step phrasing carried it.
What’s still failing?
Mostly multi-part tickets: it answers the first ask and drops the second. Round 4 is testing an explicit “address every question” clause.
Keep going.
Type a message…
Send
Settings
Extended thinking
Web search
Code execution
Optimize prompt while usingBeta
Quietly evolves parameters across your project
Get started

It just runs. Pick who runs it.

It’s just in and out. Hand over what you’re running and the examples you already have, then leave it alone. Five minutes later there’s a better version waiting with its score beside it. Watch every round if you like, or none of them. No schemas, no special code, nothing to wire up.

We run it

Online, limited

10 campaigns · up to 10 rounds each

Includes:

  • Runs on our machines — zero setup
  • Full live monitoring as it runs
  • Bring your own API key when free runs end
  • Take your results with you anytime
Open the app How it works
You run it

Secure, Personal, Local

Unlimited campaigns & rounds

Includes:

  • Claude-operated: the /potter-run skill drives it
  • Runs locally in your editor, on your own keys
  • Full control over run speed
  • The full source, nothing held back
Install guide How it works
Your team runs it

Team online

Unlimited · multi-user

Everything above, plus:

  • Multi-user from day one
  • Behind a Cloudflare tunnel + OIDC allowlist
  • Whitelabel — your brand, your invite list
Self-host guide View the source
Compare all the ways to run it →

Rewording it until it feels better
is not the same as knowing.

Every version PromptPotter writes is scored against your own examples before it is kept or thrown away. So what comes back is not the one that reads best. It is the one that measured best, and it keeps going until the score stops moving.

Try it — it’s free