Rescriptum

Native agent IDE

The self-verifying native agent IDE.

An agent that proves its work. Every edit is checked against your real build and tests before the turn can finish, and the summary is checked against git. Bring any model: OpenRouter, your own provider, or a local one.

Native DesktopWindows 10 / 11 (x64) · v0.8.0 LiveCode · Plan · Ask

Download Windows ClientDocumentation
18.4K · $0.08
routes.ts
Code
routes.ts
import { rateLimit } from "../middleware/rateLimit";
 
export async function handleRequest(req: Request) {
const limited = await rateLimit(req, { windowMs: 60_000, max: 100 });
if (!limited.ok) {
return new Response("Too many requests", { status: 429 });
}
 
const session = await authenticate(req);
if (!session) {
return new Response("Unauthorized", { status: 401 });
}
 
return routeHandler(req, session);
}
patch_file · src/api/routes.ts

Interactive preview — explore sidebar, files, and chat like the desktop app.

Core Capabilities

Three pillars of Rescriptum

Rescriptum isn't just a chat box in an editor. It's a complete agent workspace with verification, memory, and cost discipline built in.

01

Build gates & instant rollback

Verification & Safety

Every round of edits is checked by language-server diagnostics and your project's real build and tests. The agent cannot finish a turn while they fail, and its final summary is compared with git so claims have to match reality. Any turn can be rewound from a checkpoint.

  • Build and test verification, with claims checked against git
  • Checkpoint rewind and isolated git worktrees
  • Strict, Balanced, or Permissive approvals with a hard-deny policy

02

Context that compounds

Codebase Intelligence

A local symbol index lets the agent read one function instead of a whole file and find every caller. Per-project memory, a durable scratchpad, and cited research from your own browser keep it accurate across long sessions.

  • Symbol search for TypeScript, JavaScript, Rust, Python & Go
  • Cross-session project memory & learned conventions
  • Cited research and live page inspection via Rescriptum Tab

03

Bring your own provider

Your models, your costs

Connect OpenRouter, any OpenAI-compatible provider, or a local model, and mix them. Rescriptum keeps prompts cache-friendly, so cache hit rates typically land around 90 to 95 percent, and shows every token and cost in a live Treasury with a daily spend limit.

  • Native Windows app with a built-in terminal and live preview
  • Live token Treasury, cache hit rate & daily spend limit
  • OpenRouter, any OpenAI-compatible endpoint, or local models
Bring your own key. Inference is billed by the provider you choose, never by Rescriptum.

Workflow

Three modes, one engine

Ask when you want answers. Plan when you want to agree on an approach first. Code when you want the change made and checked.

Code mode

Full agent, checked at every round

Read, patch, run commands, query diagnostics, and verify against your real build and tests. The turn cannot finish while checks are failing, and the final summary is compared with git. This is the default.

Tools available
read_filepatch_filerun_commandget_diagnosticsverify+ about 50 more, plus MCP tools

Agent observability

Edit, verify, log. Recovery when tools fail.

Diagnostics run right after each write, and the build or tests run after each round. The agent sees real compiler output and has to fix it: turn_complete is rejected while checks fail, and the summary is compared with git before you see it.

Happy path
patch → diagnostics → test → turn_complete
round 1…
rollout.jsonl · debounced verifyillustrative trace
Self-correction
patch → error → read → fix → verify
round 1…
rollout.jsonl · debounced verifyillustrative trace

Verify gates

Language-server diagnostics, then your project's typecheck, build, or tests, using the toolchain of the package that was edited. Debounced so rapid patches don't spam builds.

Turn completion

The agent signals done with turn_complete. The gate rejects it while the build or checks fail, and the summary is checked against git. A stuck turn ends with the failure reported, not hidden.

Rollout logs

Full JSONL trace per turn: every tool call, cost, and round for debugging and audit.

Rescriptum Tab

Give the agent a browser it can read

A browser extension linked to the IDE over a local connection. Chrome, Edge, and Brave.

See how it works

Cited research

Ask a question and Tab searches, reads the top pages, and answers with numbered sources. The agent can do the same mid-task.

Pin what you see

Pick an element on the page. The agent gets its selector, styles, and, on React dev builds, the component and source line.

Check the real page

Read console and network errors, click through the UI, and reload to confirm a fix, all inside your own browser.

Cloud dashboard

Usage synced from the IDE

Sign in with the same account in the desktop app. Costs, models, and workspace activity update after agent turns — typically within ~60 seconds.

  • Token and cost breakdown by model
  • Per-workspace activity from usage logs
  • Device heartbeat when signed into desktop
Open dashboard
Example layoutlive when signed in
Tokens
2.4M
illustrative
Cost
$4.12
illustrative
This week
Model mix
deepseek-v4-flash 42%claude-sonnet 28%gpt-5.5 18%others 12%

Trust & safety

Serious agents need serious guardrails

Autonomy without oversight is how repos get deleted. Rescriptum layers policy, approvals, and isolation so you can delegate real work.

Terminal security

Catastrophic commands (rm -rf /, $HOME, .) are always denied. Force-push, recursive delete, destructive SQL, and curl-to-shell always ask. Every command is classified before it runs.

Tool permission presets

Strict, Balanced, or Permissive in Settings. Auto-approve is off by default, and trust only ever relaxes low-risk tools. Shell, delete, and push never auto-approve by trust.

Worktree sandboxing

Agent edits on isolated branches. Review diffs, run verify gates, merge or discard. Main stays safe.

Transparent rollout logs

Every tool call, round, and cost in JSONL. Debug agent behavior without guessing what the model did.

Use cases

Built for real repo work

Multi-round agent sessions on production codebases, not autocomplete demos.

01
Debug

Fix failing tests without a 27-round spiral

prompt › Tests fail on auth middleware. Find root cause and fix.

result › 3 rounds · patch + diagnostics · tests pass · turn_complete

02
UI

Redesign a dashboard with live preview probes

prompt › Make the analytics page responsive. Match the design system.

result › worktree branch · UI verify · Tailwind density check

03
Scaffold

Ship a feature from a Plan → Code workflow

prompt › Plan a webhook ingestion API, then implement after I approve.

result › Plan mode recon · approve · Code mode · verified every round

04
Isolation

Risky refactors in git worktrees

prompt › Migrate the terminal backend to PTY. Don't touch main.

result › agent worktree · review diff · merge or discard

05
Browser

Fix the exact element you are looking at

prompt › Pin the pricing card in the browser and match its padding to the others.

result › Rescriptum Tab pin · component + file located · edit · page re-checked

06
Research

Look it up with sources, without leaving the IDE

prompt › Why does this effect run twice in dev? Cite your sources.

result › ask_tab · 3 pages read · cited answer [1][2][3]

Live Browser CompanionPersistent Project MemoryVisual UI Verification

Platform

What ships in Rescriptum

Native agent infrastructure for serious repo work — not a chat overlay on an editor.

Verification & Safety

  • —Automatic diagnostics & build checks after each edit
  • —Checkpoint rewind to before any turn, plus git-checked summaries
  • —Isolated command execution with directory guardrails

Codebase Intelligence

  • —Symbol search across TypeScript, JavaScript, Python, Rust & Go
  • —Tolerant patching with per-file edit queues
  • —Local index and project profile built on workspace open

Multi-Agent Workflows

  • —Subagent delegation (editor, reviewer, researcher, thinker)
  • —Isolated git worktrees for safe parallel exploration
  • —Strict, Balanced, or Permissive approvals and a hard-deny policy

Desktop & Browser Ecosystem

  • —Fast native desktop client with instant startup
  • —Rescriptum Tab: cited research, element pinning, live page QA
  • —Bring-your-own providers, MCP servers, skills, and plugins
Any model
Bring your own provider
OpenRouter, OpenAI-compatible, local
90–95%
Typical cache hit rate
Varies by project, shown live
0+
Built-in agent tools
Plus MCP servers

Pricing

Simple, transparent pricing

Start free with your own key. Pro adds cloud features when billing ships — the IDE stays fully usable on Free.

Beta

Free

$0/mo
  • Full IDE: Code, Plan and Ask
  • Bring your own provider key
  • Local treasury & rollout logs
  • Cloud usage dashboard when signed in
Coming soon

Pro

$19/mo
  • Extended cloud analytics
  • Priority model routing
  • Cross-device preferences sync
  • Email support
Notify me

Enterprise

Custom
  • Self-hosted deployment
  • Private gateways and custom deployment
  • SSO & team management
  • Dedicated support & SLA
Contact sales

Download

Start with the desktop IDE

Rescriptum for Windows is available now. Verified edits, a local symbol index, and the model provider of your choice, all running on your machine.

Download for Windows (x64)
v0.8.0 Live Release · Free Setup & PortableCloud dashboard

Native Windows 10/11 · Verified edits · Local symbol index · Bring your own provider