For the post-code developer
Code is dead.
So stop reading it. Design what you want in plain language; a team of agents plans it, builds it, and checks its own work. You judge the result by running it, never by opening a file.
Proof: Panwright was built by the workflow it ships. A quarter-million lines of Rust and TypeScript, zero written by hand, in under a month. Think it’s slop? Run it.
30-day free trial
macOS, Windows & Linux/Powered by OpenRouter or your connected AI plans
New
Your AI subscriptions
Use all your subscriptions in one place, all at once.
Connect the AI plans you already pay for. Their models join one picker, so every agent role and every chat can use the plan that fits it.
- OpenAIChatGPT plan
- xAISuperGrok plan
- MoonshotKimi Code plan
- MiniMaxCoding Plan
- Run them side by side. Put the planner on ChatGPT, the leader on SuperGrok, and the builders on Kimi or MiniMax. Every role binds independently, and any chat can switch plans without changing the rest of the team.
- Honest about what it costs. Turns covered by a plan count as $0. If ChatGPT or SuperGrok draws on paid extra credits, Panwright reports the spend. Kimi and MiniMax stop when their allowance runs out and show you when it resets.
- Your accounts, your machine. You connect to each provider directly. The credentials are stored in your operating system’s keychain and never pass through our servers.
You still need an OpenRouter key, because web search and project memory run on it. That part is cheap: $50 in credits can last months. See setup.
Token costs
We do better when you spend less.
Panwright makes nothing on your token usage. OpenRouter usage is billed directly, at cost, on your own key; subscription-backed turns are covered by the plans you connect. Our income is the flat Panwright subscription. So the only way we win users is to keep making the harness do more with fewer of your tokens. Most other harnesses are sold by companies that profit directly from your usage. They have no incentive to bring your costs down. We have no other incentive.
Why we built it
We spent six months inside single-chat coding assistants, 10+ hours a day. They are brilliant. They also forget what you tell them, lose the thread on long work, and either interrupt you for every small call or charge ahead for hours making calls that were yours. Panwright is what we built to fix that.
- One memory, kept by the harness. Panwright curates a single project memory instead of letting every agent keep its own notes, and surfaces only what matters for the task at hand. Explain a decision once, and every agent knows it.
- Compaction that holds. Long runs keep their thread instead of drifting as the context fills up.
- Structured workflows and audits. Work moves through real phases, and the result gets audited, not just generated.
- Agents that talk to each other. They coordinate directly instead of routing every message back through you.
- A hierarchy that shields you. Supervisors handle the routine and escalate only the calls that genuinely need you.
- Costs: Not an arm and a leg. A spread of models at different costs and strengths puts the expensive ones only where their judgment earns it.
How it works
Three moves. No code in sight.
You design, a studio of agents builds, you verify by running it. That is the whole loop.
Design
Sketch the features and what each should do. Panwright fills in the details and flags the calls you never made, so you weigh in only where it matters.
Build
One agent plans the work and hands it out. The rest build in parallel, review each other, and test as they go. Every project remembers your decisions.
Run
No diffs to approve. Panwright shows the behaviors you asked for, working, and says in plain words what it did. Off the mark? Say what to change.
A studio, not a single bot
Pan is a general assistant that keeps you posted on your projects and steers them on your guidance. Under it, a studio of agents plans the work, builds in parallel, and reviews itself, the way a real team would. Step in anywhere, from your desktop or your phone.
Put your best model where it matters
Panwright runs on OpenRouter. Connect one key and every model on it is yours to assign. Give planning and review a frontier model, where judgment earns its cost. Hand the grunt work to a fast, cheap one. Change your mind whenever. Connect your AI subscriptions and their models join the same picker, on the plans you already pay for.
Know where every dollar goes
Panwright meters every call and folds it into a live breakdown: by project, by feature, by model, and by agent role. Set a spending cap and it pauses a project before it runs past it.
On your terms
Your machine. Your key. Your call.
Panwright runs on your computer and answers to you, not to us.
Runs on your machine
Your projects and keys stay on your device. Your work never passes through our servers.
Billed at cost
Model usage runs through your own key at cost. We add nothing on top.
Steer from your phone
Pair your phone to follow the team and redirect it from anywhere. The link is end-to-end encrypted, so our relay cannot read your data.
Setup
Connect your accounts.
OpenRouter is required and takes about a minute to set up. Subscriptions to ChatGPT, SuperGrok, Kimi Code, and MiniMax Coding Plan are optional, and add their plan-backed models alongside it.
OpenRouter Required
-
1
Create an OpenRouter account
Free to make. It holds your models and your credits.
-
2
Connect OpenRouter
Panwright opens your browser, you approve, and the key it creates goes straight into your machine’s keychain. Nothing to copy or paste. Already have a key? Paste that instead.
-
3
Add credits on OpenRouter
This pays for OpenRouter-backed agent turns. Roles bound to a connected subscription use that plan instead, so your OpenRouter credits cover only the roles left on OpenRouter, plus memory and web search.
-
4
Assign a model to each role
A frontier model where judgment matters, a fast cheap one for the grunt work.
-
5
Start building
Describe what you want and let the studio run.
AI subscriptions Optional
-
1
Connect your subscriptions
In Settings under Models, connect ChatGPT, SuperGrok, Kimi Code, MiniMax Coding Plan, or all four. You authenticate with each provider directly.
-
2
Bind each role to a plan model
Every connected plan gets its own heading in the same model picker. Mix their models with one another and with OpenRouter, per role or per chat.
-
3
Watch every allowance
Panwright shows each plan’s usage window and reset time, plus purchased or Extra Usage Credits where the provider offers them.
Pricing
Two plans. 30 days free.
The full desktop studio for $15, or add the phone companion for $30. Both start with a 30-day free trial. Cancel anytime.
or $144/year (save 20%)
Desktop. Starts with a 30-day free trial. The $15 covers the app itself. Your agents’ model usage is billed separately: at cost by OpenRouter on your own key, or covered by any AI plans you connect.
- The full Panwright desktop app
- Unlimited projects and agent sessions
- Design canvas, agent teams, workflows, and memory
- Runs on your own OpenRouter key
- Any role can run on a connected AI subscription instead
Billing is handled by our reseller Paddle. Cancel before the trial ends and you will not be charged.
Adds the phone companion
or $288/year (save 20%)
Everything in Desktop, plus your phone. Also starts with a 30-day free trial; model usage is still billed outside the subscription.
- Everything in the Desktop plan
- Pair your phone and steer from anywhere
- End-to-end encrypted, so our relay cannot read your data
The phone apps are in store review. Message us on Discord if you’d like to test the phone companion while it’s in review.
Stop reading code.
Start directing the thing that writes it. Available now for macOS, Windows, and Linux.
Download