Opencodex

Opencodex brings a Codex-style agent loop to the editor you already use. Open a folder, choose a free model, and describe the outcome you want. The agent can inspect the repository, plan the work, edit files, run commands, and show every step as it happens.

Contents

• Why Opencodex
• Quick start
• What it can do
• Model providers
• Safety and approvals
• MCP, subagents, terminals, and memory
• Skills marketplace
• Settings
• Commands
• Development

Why Opencodex

• Free to start: OpenCode works anonymously out of the box. No Opencodex account or API key is required.
• Bring your own provider: Switch between OpenCode, OpenRouter, Groq, Gemini, Mistral, and Ollama.
• Local models: Use compatible models installed in Ollama without sending prompts to a cloud model provider.
• Visible agent work: Follow plans, reasoning, tool activity, retries, and streamed answers from the sidebar.
• Workspace-aware: Each folder gets its own project and conversation history.
• Built for recovery: Editor undo and Git restore points make agent changes easier to inspect and roll back.
• Extensible with skills: Discover, preview, and install SKILL.md packages from SkillsMP or GitHub.
• Your choice of control: Confirm every action, auto-approve edits, or enable Open access.

Quick start

1. Install Opencodex from the VS Code Marketplace.
2. Open a folder in VS Code.
3. Click the O button in the editor title bar, or run Opencodex: Open Chat from the Command Palette.
4. Complete the one-time setup. The default OpenCode provider needs no account or API key.
5. Pick a free model below the composer and describe what you want to build.

That is enough to begin. API keys, Ollama, SkillsMP, and SearXNG are optional.

What it can do

Work like a coding agent

• Read and search the active workspace
• Create, edit, and delete files within the workspace boundary
• Review proposed file diffs before they are applied in Ask mode
• Run commands with configurable approvals
• Keep named shell sessions alive for dev servers and interactive follow-up commands
• Delegate bounded exploration, review, and implementation work to subagents
• Use tools exposed by configured MCP servers
• Present a pinned plan before beginning multi-step work
• Stream answers while showing collapsible tool and reasoning details
• Schedule one follow-up prompt and steer an active run
• Retry interrupted model streams instead of silently claiming completion

Keep projects organized

• Maintain separate chat history for every VS Code folder
• Switch, archive, restore, and permanently delete conversations
• Copy messages and review timestamps
• Restore Git-tracked files to the state captured after an agent response
• Receive native completion, approval, and failure notifications
• Review token usage by model and provider
• Keep durable project decisions in .opencodex/memory.md

Build and verify larger changes

• Parallel worktrees: Create isolated branches under .opencodex/worktrees, open each task in its own VS Code window, inspect its status, merge it into the main worktree, or discard it.
• Browser testing: Launch Chromium, follow the animated agent cursor in a small preview attached to its reasoning, open that preview in a read-only viewer, click controls, type through real keyboard events, press keys and shortcuts, inspect the DOM and accessibility tree, collect console and network failures, test mobile through desktop geometry, and save full-page screenshots.
• GitHub workflows: Browse issues, create an implementation branch, inspect the current pull request, and monitor CI. The agent can also create an approved pull request through the authenticated gh CLI.
• Shadow checkpoints: Capture tracked and untracked files after mutating agent tools. Restore the full workspace or selected files without changing the normal Git index. A safety checkpoint is created before every restore.
• Repository intelligence: Search symbols and semantically relevant files, or build an import and package architecture map.
• Custom agents: Define repository agents in .opencodex/agents.json, including their prompt, model, provider, allowed tools, skills, MCP servers, and routing keywords.
• Tasks and hooks: Define recurring or one-time prompts and lifecycle policies in .opencodex/automations.json. Hooks can log, warn, require approval, or block matching tool actions.

Model providers

Every provider uses the same agent experience. Keys are stored separately for each provider in VS Code SecretStorage and are never written to settings.json.

As of 0.2.8, Opencodex supports adding custom OpenAI-compatible providers. Don't see your provider on the list? Just add it in settings!

Cloud free tiers have their own rate limits and availability can change without an Opencodex release. Reopen Settings to refresh the model list. Opencodex filters out non-chat and non-tool-capable models where provider metadata allows it.

Safety and approvals

Opencodex constrains its built-in file tools to the active workspace. Common credential files such as .env are blocked, and provider API keys are kept in VS Code's encrypted SecretStorage.

In Ask mode, Opencodex opens a VS Code diff containing the proposed content before asking you to apply or reject it. Applied edits use VS Code workspace edits and participate in editor undo. Open access is intentionally powerful; use it only in workspaces where you are comfortable allowing unattended commands.

MCP, subagents, terminals, and memory

Configure MCP servers in the Settings MCP picker, sort of like ChatGPT plugins. Opencodex can use multiple servers at once, and each server can expose multiple tools. The agent can call any tool from any server, and the tool's output is streamed back to the sidebar.

Each MCP tool is namespaced by server and follows the active approval mode. Connections are opened for the agent run and closed afterward.

The agent can delegate a bounded task to an explorer, reviewer, or worker subagent. Explorer and reviewer subagents are read-only; worker subagents can edit and verify through the same approval system as the parent.

Named persistent shell sessions remain alive across tool calls and conversations until they are explicitly stopped or the extension closes. Project memory is stored in .opencodex/memory.md, loaded into future requests, and editable with Opencodex: Open Project Memory from the Command Palette.

Skills marketplace

Opencodex can extend itself with SKILL.md packages from SkillsMP or a GitHub repository.

• Discover: Search popular or recently updated skills from the shop button beside Settings.
• Preview: Read a skill's instructions and inspect its GitHub source before installation.
• Install: Add the selected skill to Opencodex's global extension storage after approval.
• Use anywhere: Installed skills are available across workspaces and added to the agent's instructions from the next request.

You can also ask directly: “Find me a skill for web scraping” or “Install the planning skill from owner/repository.”

Settings

Open the gear button in the Opencodex sidebar or run Opencodex: Open Settings.

Optional web search

Set a SearXNG base URL to enable the web_search tool. The server must allow JSON output with format=json.

• Local VS Code: http://localhost:8888
• Dev Containers or remote workspaces: use an address reachable from the extension host, such as the host LAN address or host.docker.internal where supported

See the SearXNG installation documentation for setup options.

Commands

Requirements

• VS Code 1.106.0 or newer
• An internet connection for cloud providers
• Optional: a free provider API key for OpenRouter, Groq, Gemini, or Mistral
• Optional: Ollama for fully local models
• Optional: SearXNG for web search
• Optional: Google Chrome or a Playwright Chromium installation for browser testing
• Optional: GitHub CLI authenticated with gh auth login for GitHub workflows

Development

git clone https://github.com/matonhp5108/Opencodex.git
cd Opencodex
npm install
npm run check
npm test
npm run build

Open the folder in VS Code and press F5 to launch an Extension Development Host.

npm run package

The package command type-checks the project, builds the extension, and creates a .vsix with vsce.

Known limitations

• Free-model catalogs and rate limits are controlled by each provider.
• Mistral does not expose a free-model marker, so its built-in list is curated.
• Web search requires a user-provided SearXNG instance.
• Skills search and installation use SkillsMP and GitHub; unauthenticated GitHub API limits may apply.
• Legacy response restore points cover Git-tracked files. Shadow checkpoints also include untracked files.
• Browser tests verify behavior, accessibility data, console and network health, and measurable layout. Text-only models cannot verify subjective appearance.
• Persistent terminal sessions use piped shells rather than a full PTY, so full-screen terminal applications are not supported.

Support and contributing

Bug reports and feature requests are welcome in GitHub Issues. Include the Opencodex version, VS Code version, provider, and selected model when reporting model-specific problems.

Pull requests are welcome. Please run npm run check and npm run build before submitting changes.

License

MIT

  ██████╗    ██████╗   ██╗  ██╗
 ██╔═══██╗  ██╔════╝   ╚██╗██╔╝
 ██║   ██║  ██║         ╚███╔╝
 ██║   ██║  ██║         ██╔██╗
 ╚██████╔╝  ╚██████╗   ██╔╝ ██╗
  ╚═════╝    ╚═════╝   ╚═╝  ╚═╝