EnvCAD

EnvCAD is a Windows environmental CAD application for opening, editing, annotating, measuring, and exporting DXF/DWG drawings. Its AI Assistant can use either Claude Code or OpenAI Codex to perform the same strictly bounded CAD tool operations.

The desktop application works as a normal CAD editor when one or both AI providers are unavailable. Provider, model, and effort controls are populated from the installed runtimes; EnvCAD does not maintain a stale model allowlist in the renderer.

Windows installation

EnvCAD v0.2.4 targets Windows 11 x64. Install the generated Squirrel EnvCAD-0.2.4 Setup.exe, then launch EnvCAD from the Desktop shortcut or Start menu. The installer is currently unsigned, so Windows may show a SmartScreen warning.

Version 0.2.4 lets either supported AI provider inspect the actual Sheet Preview through a bounded, in-memory image generated from the same final SVG shown in the UI. It also removes selection as a prerequisite for drawing/layer work and provides cursor-based continuation and bounded operation batches so the AI can work through large selections and drawing catalogs instead of receiving one inaccessible oversized result. Inspect with AI uses the selected provider, model, and effort without switching providers or granting screen, browser, shell, or filesystem access.

At sidecar startup EnvCAD loads, integrity-checks, and compiles the bundled, pinned CAD and DXF SKILL.md sources from earthtojake/text-to-cad through EnvCAD's native compatibility layer. Every accepted turn activates the verified cad-core and dxf-core manifests before acknowledgment; intent-specific compiled fragments are added without repeatedly injecting the complete source corpus. The UI shows every activation and its integrity status. Attribution and the MIT license are in THIRD_PARTY_NOTICES.md.

Revision-bound sheet, model, region, selection, comparison, and overlay captures carry dimensions, render settings, evidence IDs, and SHA-256 digests. The renderer and sidecar independently validate image metadata and content before a provider can receive it. Claude runs one non-persistent streaming SDK session per in-app conversation, so preview Base64 is not written to Claude project transcripts; the upgrade also removes only legacy EnvCAD-runtime transcript directories created by earlier development builds. The explicit document lifecycle, deterministic Fit Drawing, per-drawing sheet setup, and unit/clipping safeguards from 0.2.2 remain intact.

The AI Assistant requires whichever local provider you intend to use:

• Claude Code 2.1.220, authenticated with an existing Claude subscription login (claude auth login).
• Codex CLI 0.145.0, authenticated with an existing ChatGPT login (codex login).

EnvCAD deliberately rejects ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN, CLAUDE_CODE_OAUTH_TOKEN, OPENAI_API_KEY, CODEX_API_KEY, and CODEX_ACCESS_TOKEN. It neither reads nor packages provider authentication files.

After launch, open AI Assistant, select a provider, choose one of its live models, and optionally choose an advertised effort. Default delegates effort selection to that provider. A prompt is not sent until the sidecar acknowledges the exact provider/model/effort configuration. Switching configuration starts a new conversation; there is no provider-to-provider fallback.

Core CAD features

• Open ASCII DXF and supported DWG files.
• Start a clean document with New Drawing, then render, pan, zoom, select, inspect layers, and use deterministic Fit Drawing.
• Draw and transform lines, polylines, rectangles, circles, arcs, text, hatches, dimensions, leaders, and environmental symbols.
• Ask the AI to inspect the whole drawing or any layer without selecting objects first. Large selections, entity catalogs, layer catalogs, long text, and long polylines expose lossless continuation pages that the agent is required to consume; large edits use internal operation batches with no total entity-count limit.
• Measure area, length, overlap, containment, and clearance using drawing database geometry.
• Import CSV boundaries and GeoJSON geometry.
• Configure page size, scale, orientation, and title blocks.
• Save searchable DXF and export vector PDF.
• Autosave, undo/redo, single-instance desktop behavior, and clean install/uninstall lifecycle.

Multi-provider architecture

The packaged renderer runs on a random loopback origin with Electron context isolation and sandboxing enabled. A token- and origin-authenticated loopback WebSocket connects it to one Electron utility process.

The utility process contains a provider-neutral coordinator:

• ClaudeProvider uses the installed Claude Agent SDK and Claude Code subscription login.
• CodexProvider uses codex app-server --stdio with the existing ChatGPT login.
• Both adapters are generated from one neutral 54-capability catalog, including bounded local-input retrieval and revision-bound vision.
• An app-owned broker re-authorizes every call against verified active skills, schemas, data scope, context capacity, and capability policy.
• Immutable reads may run concurrently. Mutations pass through one durable operation coordinator with complete workspace revisions, idempotency keys, receipts, postcondition checks, reconciliation, and AI-action undo groups.
• The providers receive no arbitrary renderer IPC.

Codex runs in an empty per-session directory under %LOCALAPPDATA%\EnvCAD\ai-runtime, with a read-only sandbox, never approvals, web/shell/apps/connectors/plugins/skills/subagents disabled, and every user-configured MCP server explicitly disabled. A short-lived zero-turn config/read probe discovers validated MCP names; a separate production process then proves those servers are disabled and inert and that no unexpected codex_apps surface exists before account/model access or conversation startup. EnvCAD validates runtime settings and passive-notification schemas, requires ChatGPT authentication to remain active, pins the official OpenAI provider and ChatGPT backend, re-attests the account before discovery and every conversation, and treats command, file, web, app, connector, MCP, or subagent events as security failures.

User prompts have no fixed character-count limit. Inline UTF-8 instructions up to 128 KiB use one protocol-v2 turn envelope; larger text and attachments are streamed in independently hashed 256 KiB chunks to a 1 GiB-quota local authoritative store. Providers receive a reference plus bounded outline, search, chunk, range, and metadata tools. A conservative context budget counts system instructions, the current prompt, tool results, and image reserves. Exact selection IDs stay frozen inside the renderer and do not get copied into the model prompt. Each AI turn and operation is bound to the complete workspace revision (document identity plus document, content, sheet, and view revisions); stale commands are rejected before CAD execution. ultra is never advertised as an effort, and a model whose provider default is ultra is omitted because EnvCAD is intentionally single-agent.

Accepted turns are journaled before acknowledgment and emit monotonic, replayable protocol-v2 events through exactly one terminal outcome. Drafts, queued follow-ups, partial output, turn cursors, and operation receipts survive reconnects and renderer reloads. In the packaged Windows app, renderer recovery state is atomically mirrored under Electron userData and encrypted with Windows protected storage so it also survives the random loopback origin changing after an application restart. The sidecar and provider supervisors use bounded restart, backoff, health, and circuit-breaker policies; EnvCAD never silently switches providers.

Claude is configured with no built-in tools and may invoke only mcp__cad__*. Its settings, skills, and plugins are excluded. Its SDK query is streaming and non-persistent (persistSession: false) and never resumes a disk-backed transcript.

See desktop architecture, agent protocol, and benchmark method.

Source development

Prerequisites:

• Windows 11
• Node.js 24 or a compatible current Node.js release
• npm
• Provider runtimes only when exercising live AI

git clone https://github.com/ljdstechva/EnvCAD.git
Set-Location EnvCAD
npm ci
npm run dev

For the Electron application:

npm run desktop:dev

The browser development server uses the deterministic fake sidecar for E2E tests. Normal automated tests never send real provider requests.

Verification

npm run test
npm run typecheck
npm run build
npm run test:e2e
npm run test:desktop
npm run desktop:make
npm audit --omit=dev
git diff --check

The explicit live benchmark is opt-in:

npm run benchmark:ai -- --live

It uses only a generated metre-unit drawing and the two deterministic benchmark prompts. Raw transcripts, DXFs, screenshots, and machine-specific timings are written under ignored output/desktop/ai-benchmark/. The command refuses to make a model call without --live and enforces its own turn budget.

The installed visual acceptance is also opt-in and requires the M-01 drawing:

npm run acceptance:visual-installed -- --live --scope=full --drawing "C:\path\to\M-01.dxf"

It sends bounded Sheet Preview images to both selected subscription providers and writes screenshots plus a machine-readable report outside the source tree. Release acceptance additionally drives both live providers through the actual installed Windows shortcut with OS-level UI automation, records the installed process path and screenshots, verifies normal close/runtime cleanup, and merges that evidence into the report.

See TESTING.md for suite boundaries and installed-app checks.

Keyboard and appearance

• Ctrl+O — open a drawing
• Ctrl+S — save DXF
• Ctrl+Z / Ctrl+Y — undo / redo
• Delete — erase selected entities
• Middle-button drag — pan
• Mouse wheel — zoom
• Theme toggle — light/dark canvas and UI

File safety

EnvCAD validates obvious invalid files before opening, keeps unrelated files untouched, and writes user-selected DXF/PDF outputs only. AI preferences contain provider/model/effort identifiers and optional benchmark recommendations—never credentials. Corrupt preferences recover to the existing-user default of Claude Code.