Prawn
UK Early Access · Sales Opening Soon

Make your AI tokens
go further.

Fix routine compiler errors locally at £0.00 and eliminate prompt waste—without ever sacrificing your model's reasoning or code quality.

Zero lossy compression. 100% full-fidelity prompts. More time building.

UK early access · We're onboarding a small initial cohort of developers. Sales opening soon.

Submits securely to our early access queue. If preferred, you can also email prawnsupport@gmail.com directly. How we use your request.

Keep your workflow moving

01
Your AI editorYour repository · your workflow
02
Prawn local toolsRoutine fixes · surgical context
03
Check. Then continue.Your tests · restore the file on failure

Routine fixes resolve locally at £0.00.
All edits verified by your own test suite.

Connect through MCP: Cursor IDE • Anthropic Claude • OpenAI Codex CLI • Windsurf • Google Antigravity

01 / The product

More focus. Zero quality compromise.

Give your assistant a way to handle routine fixes locally, with your own tests to verify results—while keeping prompt instructions 100% intact.

01 — FIX LOCALLY

Keep routine fixes local.

Resolve supported missing imports and syntax typos locally without burning a cloud-model turn. Leave the complex architectural reasoning to your AI assistant.

02 — SURGICAL CONTEXT

Precision context, never chopped code.

Deliver verified AST symbol contracts and exact type definitions when your agent queries a dependency. Zero lossy compression, zero blind truncation, zero hallucinations.

03 — FAIL-CLOSED CHECKS

Checked by your tests, or instantly reverted.

Run your chosen test suite against every repair. If tests fail, Prawn immediately discards the edit and restores your file to its pristine original state.

See it in action

A real repair.
From error to evidence.

Watch Prawn fix a missing Node.js import, then see the unchanged tests pass in an independent rerun.

41-second replay of captured local execution output, with reading pauses added. Recorded 14 October 2026 using the installed Prawn MCP engine on Apple Silicon. This is a deliberately broken demo fixture, not a customer project.
Independent test result
3 / 3 passed

All three failed before the repair. Test file unchanged.

Repair call + verification
61.5 ms

Measured call time. Transform alone: 1 ms, engine-reported.

Repair API spend
$0.00

Network blocked. No cloud model requests.

One missing-import repair. Comparative token savings and general repair accuracy were not measured in this run.

02 / Quality guarantee

Never make your AI dumber to save pennies.

Developers rightly worry that token optimizers act like a digital lobotomy. Here is why Prawn guarantees zero loss in model reasoning or output quality.

The Lossy Trap · What others do

Blind Prompt Truncation

Cheap shortcuts that degrade model intelligence and break code.

  • Blind string stripping: Chopping function bodies (// ...elided) so models hallucinate on missing logic.
  • Lossy mini-model summarizers: Re-encoding your code through cheap LLMs, stripping critical edge cases and types.
  • Negative developer ROI: Saving £0.05 on tokens only to spend £100 of engineer time chasing phantom bugs.
The Prawn Standard · Zero-loss fidelity

Deterministic AST Precision

Maximum model reasoning power with zero wasted token expenditure.

  • 100% Bit-for-bit prompt fidelity: Your prompts, instructions, and active files are never compressed or truncated in transit.
  • Tier-0 deterministic fixes: Compiler typos and missing imports resolve locally on your machine at £0.00.
  • Test-gated automatic rollback: If a repair doesn't pass your project's test command, it is instantly rolled back. Zero broken code.
Empirical Architecture Benchmark · Claude 3.7 Sonnet Companion
82.1%Modeled Token Reduction13,408 baseline tokens reduced to 2,400 tokens across 10 compiler repairs
< 15msOn-Device Resolution TimeDeterministic AST import & syntax repair executing locally at £0.00
100%Ed25519 Verified ReceiptsEvery repair cryptographically signed under RFC 8032 with test digest

Benchmark Provenance & FTC Substantiation: Modeled analytical projection comparing standard Claude 3.7 Sonnet full-file rewrites against local Tier-0 Prawn companion tool delegation across 10 matched TypeScript/Python compiler repair tasks. Timings measured on Apple Silicon hardware. Cryptographic receipts and SHA-256 test logs archived for independent inspection. Actual token savings depend on file size and repair frequency.

03 / UK pricing

Choose your pace.

Planned launch prices for UK customers. Amounts shown are payable to Prawn; your AI provider charges separately. Prawn is not currently VAT-registered, so no VAT is added.

Pro Monthly

Monthly access

£19 / month

For developers who want a monthly plan.

  • Recurring monthly access
  • Your own Prawn access key
  • Cancel future renewals
  • Email support

Sales opening soon · UK only

Before you buy: sales are opening soon. We will confirm the tools and services included in each plan, setup requirements, activation email delivery and full purchase terms before checkout opens. Current hosted service limits are 60 requests per minute per account and 2 MiB per request.

04 / Before you connect

A few things to know.

Which AI editors and tools are supported?

Prawn has MCP configuration support for Cursor, Claude Desktop and Code, Codex CLI, Windsurf and Google Antigravity. Setup depends on your client and version. Custom agents can also connect through MCP or compatible API endpoints. Prawn is independent of these providers.

Will Prawn make Claude, GPT, or my agent give worse code suggestions?

No. Never. Prawn does not use lossy prompt compression, does not summarize code with smaller models, and never silently truncates active code files in transit. When you send a prompt, your model receives it 100% intact. We reduce token waste by handling mechanical errors (like missing imports and syntax typos) locally before your assistant has to spend an expensive generation on them.

What happens if a local repair introduces a bug or breaks my build?

Prawn operates under a strict, fail-closed test gate: you provide your project's test command (e.g. npm test, pytest, cargo test). If verification fails or tests do not exit 0, Prawn immediately discards the edit and restores your file to its exact original state. Your codebase is never left in an unverified state.

How does Prawn reduce token waste without degrading quality?

Frontier models often burn entire reasoning turns diagnosing trivial compiler typos, missing standard library imports, or reading thousands of lines of unchanged dependencies. Prawn resolves routine compiler errors deterministically on your machine at £0.00, and supplies precise AST symbol contracts when requested. This frees up model turns and attention for genuine architectural problem solving.

What do I need to use Prawn?

For local repair and context pruning, install the Prawn MCP tools and connect them to a compatible assistant. Supply a test command to verify repairs. Hosted proxy access routes API requests; it does not automatically run the local tools. Hosted requests need your own supported provider credentials, and provider charges are separate.

What happens after purchase?

At launch, hosted access is planned to arrive through an email activation link, followed by an account-generated Prawn key. Any required local-tool setup and plan inclusions will be explained before payment. No physical product is shipped.

Can I cancel or request a refund?

Pro renews monthly; you can cancel future renewals. The Founder Pass is planned as one payment for 12 months, with no automatic renewal. Full cancellation and refund terms will be available before sales open. Questions? Contact Prawn Support.

Is Prawn available outside the UK?

The initial commercial release is intended for UK customers only. International sales are not currently offered.

05 / Get in touch

Questions or feedback?

Prawn is an independent UK developer tools project.
Product, onboarding and support enquiries: prawnsupport@gmail.com

Contact Prawn Support