Xratu

English · فارسی

Open-source AI coding agent for VS Code. Bring your own key or use a local runtime. Everything runs inside the extension host, on your machine.

Your keys. Your models. Your machine.

API keys live in VS Code secrets storage and go only to the endpoint you configure.
Fully offline capable: Ollama, LM Studio, vLLM, and llama.cpp are auto-discovered on your machine.

Features

Agentic coding - reads files, edits code, runs terminal commands, searches the web. Every edit and command needs your approval; per-tool auto-approve if you trust a workflow, YOLO mode if you don't want the gate.
Plan mode - read-only planning with a task list. Mutating tools are blocked in code, not by prompt.
Checkpoints - shadow-git snapshots before sends and edits, restorable anytime. Your real git repo is never touched.
MCP servers - stdio, WebSocket, or streamable HTTP/SSE, with Cline-compatible files-based config.
Agent Skills - the open agentskills.io standard; skills can be shared with Claude Code, Roo Code, OpenCode, and others.
Long sessions - steer mid-task, edit & resend any message, image and PDF attachments, automatic context compaction.

Models

Group	Providers
Cloud	OpenAI, Google Gemini, OpenRouter, xAI, Groq, DeepSeek, Mistral, Perplexity, Cohere, Together, Fireworks, Cerebras, NVIDIA NIM, Hugging Face, SambaNova, Moonshot, Z.AI, OpenCode Zen
Local	Ollama, LM Studio, vLLM, llama.cpp
Custom	Any OpenAI-compatible HTTPS endpoint

Getting started

Install from the VS Code Marketplace: xratu.xratu

Prefer a manual install? Grab the latest xratu-*.vsix from Releases and install it via Code → Extensions → ⋯ → Install from VSIX…

Open the Xratu panel from the activity bar.
Pick a preset or a local runtime, paste your API key, choose a model.
Chat. Approve edits and commands as the agent works - or turn on auto-approval once you trust the loop.

MCP servers

Xratu reads server config from two files:

Global: mcp.json in the extension's global storage
Workspace: .xratu/mcp.json in the workspace root (overrides global per server key)

Format is Cline-compatible: command/args/env for stdio, url for remote. Tools surface to the agent as mcp__<server>__<tool> and follow the same approval flow as built-ins.

Agent Skills

A skill is a folder with a SKILL.md file - YAML frontmatter (name + description) plus a markdown body. Only the name/description load up front; the agent pulls in the rest via its skill tool when your request matches. Search order (first match wins):

.xratu/skills/<name>/SKILL.md
.agents/skills/<name>/SKILL.md (shared with other agent tools)
.claude/skills/<name>/SKILL.md (Claude Code)
~/.agents/skills/<name>/SKILL.md
~/.claude/skills/<name>/SKILL.md (Claude Code)

---
name: deploy-staging
description: Deploy the app to staging; use when the user asks to deploy or ship to staging.
---

1. Run the test suite
2. Build the app
3. `npm run deploy:staging`

Extra files (scripts, references, templates) go next to the SKILL.md. Enable/disable skills in Settings → Servers & skills → Skills.

Privacy

API keys are stored in VS Code secrets storage and sent only to the endpoint you configure. Chat history and checkpoints live on your disk only.

Development

npm ci
npm run compile        # host bundle (esbuild) + webview (vite)
npm run test:webview

CI type-checks host and webview separately and runs the test suites on Ubuntu and Windows. See AGENTS.md for details.

Contributing

Issues and PRs welcome. Open an issue before large changes. Conventional Commits, CI green, and code review required for every PR.

License

Apache-2.0