ParadigmEve

Make ChatGPT useful for work beyond a single chat.
Eve connects a ChatGPT conversation to real files, local tools, durable Plans, and ongoing work on your computer—so projects survive interruptions and the assistant can help finish them.

Downloads  ·  2.4.5 release notes  ·  Changelog  ·  Setup guide

Open-source beta. Download only packages actually listed on GitHub Releases. Source branches, Git tags, installed test builds, and published installers are not interchangeable. Packages may be unsigned.

Why Eve instead of only chatgpt.com?

ChatGPT already offers memory, projects, apps, and other useful features. ParadigmEve adds a desktop execution and continuity layer for people whose work lives outside the conversation.

Your work	What Eve adds
Editing a real project	Permission-scoped local files, patches, commands, builds, and test results in the same workflow.
Working past one chat	Durable local sessions and Plans, reusable helper chats, recovery after restart, and Compact & Resume handoffs.
Repeating work	Locally backed schedules, routines, saved context, and reviewable progress instead of relying on a chat reply as proof.
Doing work on your own hardware	Local file operations, tests, and supported audio processing can run on your computer rather than invoking a separately billed cloud model for every operation.
Using desktop software	Optional, separately permissioned screen, control, and clipboard tools, plus a broker for inspecting office formats and discovering available editors.

What “unmetered local work” means: reading files, running a local test, or doing local processing does not itself consume a separately metered OpenAI API inference call. Asking ChatGPT to plan, invoke tools, analyze results, or continue a conversation still uses your actual ChatGPT account access and limits. API-based services may charge separately, and local models consume your own hardware resources. Eve does not unlock unlimited ChatGPT messages, reasoning, context windows, models, or rate-limit exemptions.

Does it feel more reliable?

In our hands-on use, GPT-5.6 Sol Medium and GPT-6 High workflows have seemed more reliable through Eve than the basic ChatGPT Plus browser experience. This is a subjective, end-to-end observation, not a controlled model benchmark. The likely advantage is Eve's recoverable work state, exact tool evidence, Plans, workers, and handoffs—not a claim that Eve makes the underlying models smarter or increases their quotas. Available models and reasoning levels remain controlled by your ChatGPT account and your own model-picker choice.

What can you do with it?

• Finish software and file work: inspect and edit approved project files, run terminal commands and tests, coordinate steps through Plans, and audit the actual results.
• Keep complicated tasks moving: delegate separate pieces to reusable helper chats, preserve the main Eve conversation as coordinator, and pick up after a restart or a long-chat handoff.
• Choose how helpers work: ChatGPT workers keep individually reviewable helper chats; Native subagents work inside an active ChatGPT Work conversation and return to its parent without a separate Eve worker chat; OpenRouter and Ollama use an explicitly configured provider/model. Native subagents have not yet been proven as an unattended Eve worker/heartbeat executor, so Eve must hold such work for a verified parent instead of silently switching modes or claiming it finished.
• Notice when ChatGPT itself needs you: the 2.4.5 Companion can detect a current-turn tool approval card even when ChatGPT has collapsed it under Previous activity, record the wait, and surface a native notification. Eve does not click or auto-approve the provider card, and disappearance alone is not proof of approval.
• Remember useful context on purpose: save Pins in Threads, create reusable prompts in the %instructions category (invoke an individual one by its exact %name), group Threads into Quilts (#name), and review locally recorded conversation evidence.
• Run local and scheduled workflows: set up routines, use a bounded read-only heartbeat to review recorded requests and Plans, and optionally enable Goal/Loop continuation. The heartbeat reviews; it does not autonomously execute discoveries.
• Transcribe and work with documents: supported Windows x64 builds can use local Whisper transcription and speaker diarization, with optional explicitly selected cloud speech recognition. DOCX, XLSX, PDF, and PPTX inspection uses the built-in file broker; editing depends on an available provider and consent. WordCraft and GridCraft are separate developing apps, not bundled full-fidelity editors.
• Help across desktop apps: explicitly enable Computer use for screen inspection, mouse/keyboard interaction, or clipboard access when needed.

None of these capabilities require you to enable every permission.

Get started

1. Choose an available package on GitHub Releases and open ParadigmEve.
2. In Settings → Workspace → Folders → Add, approve a folder for Eve to work in; review Permissions, especially Run commands.
3. In Settings → Setup, configure the connection and add/pair the browser Companion as guided.
4. At chatgpt.com/plugins, choose Add → Add custom MCP server and connect Eve; review its ChatGPT action permissions. Your account must support custom MCP apps.
5. Give Eve a concrete job. Select your preferred model when offered; Eve must not silently switch it.

See the setup reference for detailed connection and platform instructions, or 2.4.5 notes for the latest source and acceptance limitations.

Privacy, permissions, and limits

Local chat recording is on by default to support continuity, review, and auditability. You can turn it off in Workspace → Permissions or beside Rebuild archive and Open Archive Browser; Setup and Settings also point to this control. With recording off, Eve stops storing new local session messages and tool history, and recording-dependent features such as Goal/Loop are unavailable. It does not erase earlier sessions, archive copies, Pins, Plans, app logs, separate outputs, or ChatGPT's own history. Local storage is not a blanket privacy guarantee; review what you save. Archive and retention details.

Files require approved folders, and Computer use is opt-in. Run commands is enabled by default but executes under your operating-system user account, not inside a sandbox restricted to approved folders. Disable it in Workspace permissions until needed if you want a narrower starting point. Protect important files with backups and read tool receipts instead of trusting an assistant's assertion alone. Security details.

ParadigmEve requires an eligible ChatGPT account and Chrome/Chromium, Edge, or Brave for its Companion workflow. Supported desktop targets include Windows 10/11, macOS 13+, and current Linux; Windows x64 has the most live installed test coverage. Other platforms, cloud providers, GPU availability, editing tools, and account features vary.

Release status

2.4.5 source (October 28, 2026): Windows x64 c26 passed 6,232 local tests and was silently installed with a matching packaged/installed application hash. Cumulative fixes protect active Compact & Resume turns, worker ownership, archive continuity and Native heartbeat transport. One Native in-band heartbeat was completed after four exact Plan acknowledgements and a durable receipt without another browser Send; a later review was truthfully blocked on unresolved historical worker evidence. Manual hosted-child review succeeded, but idle background cognition, automatic app-owned child dispatch and repeated mode-switch queue retirement are not yet proven. c26 still embeds Electron 43.8.3 / Node 24.22.0; the security update to Electron 43.8.4 / Node 24.22.1 needs a separately verified candidate. ChatGPT tool approvals remain user-controlled. See 2.4.5 release notes for evidence and limits. A source tag or branch push does not publish installers or promote Latest; consult GitHub Releases for downloads.

Build and contribute

Start with CONTRIBUTING.md, AGENTS.md, and the engineering vault.

npm ci
npm run dev
npm run verify

ParadigmEve builds on the open-source Chat-on-Steroids project. Thanks to the contributors and testers who uncover difficult continuity and recovery cases.

Downloads · Changelog · Security · MIT License

_{ParadigmEve is not affiliated with or endorsed by OpenAI. ChatGPT and Codex are OpenAI trademarks.}