Sovereign Code Runtime
Enables a ChatGPT Web conversation to operate a Windows computer through OpenAI's Secure MCP Tunnel, with configurable permission levels, local tool execution, and audit logging.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Sovereign Code RuntimeList my recent task history with audit evidence."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Sovereign Code Runtime
Source preview: this checkout is not an announced binary release or an installed-state verification. Review security and reporting, dependency findings, and contribution guidance. Source CI does not certify native desktop behavior; the complete dependency-audit job checks the development/build graph separately from the production-only audit.
Sovereign lets a ChatGPT Web conversation operate a Windows computer through MCP. ChatGPT supplies the model and conversation; Sovereign supplies local tools, permissions, task state, and audit evidence. The installed application has no dependency on Not-Code or Not-Meta accounts or services and does not include a built-in model client.
Ready-to-use builds · Documentation · Development · Android preview
Use Sovereign
Double-click 打开成品.cmd or run pnpm products to locate existing builds; this does not install or start SO. See the release guide for selection and verification details. Source checkouts, packaged candidates, and the installed application can be different revisions; a newer file is not proof that it is running.
Start Sovereign and authorize the folder the agent may use.
Make the official Windows
tunnel-clientavailable onPATH, select it locally, or setSCR_TUNNEL_CLIENT_PATHbefore launch.In ChatGPT Connection / ChatGPT 连接, enter the OpenAI tunnel ID and runtime key, select the needed permission level, and press Start connection / 启动连接.
Wait for Ready, then configure the ChatGPT custom app to use that tunnel ID. Verify a read-only tool call before continuing the task.
The normal connection is:
ChatGPT Web → OpenAI Secure MCP Tunnel → loopback Gateway
→ permission and workspace policy → local tools and auditThe desktop Gateway stays on 127.0.0.1. The connector receives its local Bearer through its child environment, not its command line. Persisted tunnel keys and proxy settings use Windows-DPAPI-protected ciphertext; plaintext stays in trusted-process memory. Optional control-plane proxies do not proxy local MCP traffic. Connection recovery still requires a working route to the official Tunnel service.
For proxy/failover, post-sign-in remote hosting, and recovery, use Remote Host and control-plane failover. A trusted HTTPS /mcp bridge is an optional alternative; without one, Copy MCP connection targets only the local Gateway. Copied bundles contain a credential and have a bounded clipboard lifetime. Rotating that credential disconnects existing sessions and stops the active tunnel.
Related MCP server: GPT Agent
Capabilities and permissions
The workbench provides Home, Tasks, ChatGPT Connection, Task History, Terminal, Python, Browser, Desktop Control, Workflows, and Settings. It combines project/task progress with managed runs and bounded activity; completed audit receipts remain the evidence for actual outcomes.
Discover the current tools through MCP tools/list, the authenticated Gateway /v1/manifest, or capabilities.search and capabilities.describe. The live manifest is authoritative; optional packs and installed revisions change availability. See capability discovery, tool packs, and task coordination.
Profile | External tool behavior |
L1 Observe | Read-only observation tools. |
L2 Workspace | L1 plus contained workspace writes, local Git writes, fixed validation, and run cancellation. |
L3 Consequential | L1/L2 directly; consequential calls require fresh native approval. |
L4 Bypass | Declared tools without Sovereign approval prompts, after explicit local confirmation. |
Remembered permissions are workspace-bound. Restoring L4 requires a valid current-user DPAPI grant for that exact workspace. L4 does not grant administrator elevation or remove schemas, containment, revision checks, or audit. Workspace restore is a separate setting: enabling or disabling it never changes the active L1-L4 permission.
PowerShell, ConPTY, Python (including -I), browser evaluation, and desktop control run with the Windows user's rights and are not OS sandboxes. Only execution explicitly routed through the optional secure-execution pack uses its reviewed Docker boundary. Secure Desktop, credential entry, elevation, passkeys, OTPs, and CAPTCHA remain human handoffs. See the threat model.
Sovereign has no notification milestones, content-hash deduplication, or per-domain/rolling notification buckets. Notifications are direct semantic signals for meaningful work, completion, or operator attention; routine successful non-workflow runs do not require an automatic notification.
Develop
From the repository root, with Windows 10/11, WebView2, Node.js 24+, pnpm 10+, the Rust MSVC toolchain, and the .NET Framework C# compiler available:
pnpm install --frozen-lockfile
pnpm dev:desktopThis starts a development shell. For targeted validation, packaging, and authorized deployment, follow development and release validation. Choose checks for the changed behavior; the aggregate commands already include overlapping tests and build steps.
Location | Responsibility |
| Primary Tauri/WebView2 shell, Rust authority boundary, packaging, and Guardian. |
| Shared workbench UI used by both Tauri and Electron. The rest of |
| Node sidecar hosting the shared control plane and Gateway. |
| MCP transport, live catalog, and optional headless entry point. |
| Settings, permissions, tasks, tunnel supervision, and runtime lifecycle. |
| Policy and ledgers; Windows primitives; tool definitions and dispatch. |
| Independent Gradle Android preview; see its README for this checkout's actual status. |
Architecture explains the execution paths. AGENTS.md governs repository work; load a topic's detailed guide only when that task needs it.
Release evidence and boundaries
Build success, signed-package verification, installation, activation, and observed running behavior are separate results. Preserve user settings and databases during deployment. A dry-run, renderer smoke, or portable smoke does not prove an NSIS installation succeeded. Installation/restart smoke replaces or stops applications and requires authorization for those effects.
Renderer updates, Runtime Host updates, and full-application restarts have different compatibility and coordination rules. Use the applicable renderer, Runtime Host, or restart guide before changing running state. Signed-update support requires provisioned trust material; it does not imply an enabled downloader or trusted release feed. Development executables are not Authenticode-signed.
The project is a clean-room implementation: it does not copy proprietary Not-Code implementation, private schemas, credentials, user data, branding, or service logic.
License
Sovereign's original source code and documentation are licensed under the Apache License, Version 2.0. Copyright 2026 Sovereign Code Runtime contributors. Third-party components retain their own licenses; see third-party notices. This license does not grant rights to third-party services, credentials, or trademarks.
This server cannot be deployed
Maintenance
Related MCP Connectors
Your org's AI agents, tasks, runs, search, and brain files as MCP tools and resources.
Authenticated MCP Agent (Openai)
MCP connector that lets ChatGPT list, search, and run your Apple Shortcuts via a local Mac agent
OCR, transcription, file extraction, and image generation for AI agents via MCP.
Related MCP Servers
- AlicenseNot gradedqualityAmaintenanceEnables ChatGPT to access approved local Windows folders, run commands, control the desktop, and manage session history through a permission-bounded MCP server.3,972MIT
- AlicenseNot gradedqualityAmaintenanceEnables ChatGPT on Windows to securely access codebases, Git, terminals, language servers, debuggers, SQLite, local HTTP services, adaptive project memory, engineering skills, checkpoints, audit logs, and sandboxed command execution through MCP tools.MIT
- FlicenseNot gradedqualityBmaintenanceEnables AI agents to control and interact with a local Windows computer through authenticated MCP tools, including file operations, shell commands, process management, HTTP requests, and system/media inspection.-
- AlicenseNot gradedqualityBmaintenanceEnables MCP-capable AI assistants to control a Windows computer through GUI automation, file and process operations, PowerShell commands, durable background jobs, and binary file transfer with downloadable exports.MIT