SideCar — AI Coding Assistant for VS Code

SideCar is a free, self-hosted VS Code extension that serves as a drop-in replacement for GitHub Copilot and Claude Code. Use local Ollama models, the Anthropic API, LLMManager, or any OpenAI-compatible server (LM Studio, vLLM, llama.cpp, OpenRouter) for AI-powered coding — with full agentic capabilities, inline completions, and tool use.

A free, open-source, local-first autonomous AI agent for coding. Full agent loop, not just chat. No subscriptions, no data leaving your machine.

SideCar will always be free, tips not required but appreciated.

Why SideCar?

Most local AI extensions for VS Code are chat wrappers or autocomplete plugins. SideCar is a full agentic coding assistant — closer to Claude Code or Cursor than to a chatbot.

What sets SideCar apart

• True agentic autonomy — SideCar doesn't just answer questions. It reads your code, edits files, runs tests, reads the errors, and iterates until the task is done. Other local AI extensions stop at chat and autocomplete.
• No vendor lock-in — Use Ollama for fully offline operation, the Anthropic API for Claude, or any OpenAI-compatible server (LM Studio, vLLM, llama.cpp, OpenRouter). Same interface, your choice.
• Security from the ground up — Built-in secrets detection and vulnerability scanning run automatically after every file write. No other local-first extension does this.
• Extensible with MCP — Connect external tools (databases, APIs, custom scripts) via the Model Context Protocol. SideCar treats them as first-class tools alongside its built-in ones.
• Production-grade safety — Agent mode controls (cautious/autonomous/manual), iteration limits, token budgets, cycle detection, diff preview, and one-click rollback keep you in control.
• .sidecar/ project directory — Persistent project storage for workspace index cache, agent reasoning logs, session history, plans, and memory. Committed project instructions live alongside gitignored ephemeral data.
• Smart context — AST-based extraction for JS/TS, Python, Rust, Go, and Java/Kotlin files. SideCar sends relevant functions and classes to the model, not entire files.

Features

Agentic Coding Assistant

• Tool use — the model can read, write, edit, and search files, run commands, check diagnostics, and run tests autonomously
• Agent loop — multi-step execution (read code, edit, run tests, fix errors) without manual intervention at each step
• Agent progress — live step count, elapsed time, and token usage during agent runs
• Activity indicators — animated progress bar and tool execution pulses so you always know SideCar is working
• Stop button — abort the agent loop at any time (Send button toggles to Stop while processing)
• Diagnostics integration — reads compiler errors and warnings from VS Code's language services
• Test-driven loop — runs tests, feeds failures back to the model, iterates until passing
• Undo/rollback — revert all AI-made file changes with one click
• Diff preview — in cautious mode, file writes open VS Code's diff editor for review before applying
• Safety guardrails — agent mode dropdown (cautious/autonomous/manual) in the header, iteration limits, token budget
• Thinking/reasoning — collapsible reasoning blocks from models that support extended thinking (Anthropic) or <think> tags (qwen3, deepseek-r1)
• Verbose mode — /verbose to show system prompt, per-iteration summaries, and tool selection context during agent runs
• Smart context selection — AST-based parsing extracts relevant functions, classes, and imports from JS/TS files instead of including whole files in context
• Bounded caches — workspace file content and AST caches use TTL-based eviction to prevent unbounded memory growth during long sessions
• Persistent shell — run_command uses a long-lived shell process; env vars, cwd, and aliases persist between calls. Supports configurable timeouts, background commands, and streaming output
• Context pruning — conversation history is automatically compressed between turns so local models don't choke on accumulated context from prior tool calls
• Clean tool display — tool calls show as Read src/foo.ts with icons and spinners, matching the polish of Claude Code and Copilot

Inline Chat (Cmd+I)

• Edit code in place within the editor
• Select code and describe changes, or insert at cursor
• Uses surrounding code context for better edits

Inline Completions

• Copilot-like autocomplete as you type (opt-in via settings)
• Uses Ollama's FIM endpoint for local models
• Falls back to Messages API for Anthropic
• Debounced with in-flight cancellation

Code Actions

• Right-click menu: Explain, Fix, Refactor with SideCar
• Selected code is sent to the chat with the action

AI Chat

• Streaming responses in a dedicated sidebar panel
• Workspace indexing — persistent file index with relevance scoring replaces per-message glob scan, updated incrementally via file watcher
• Rich markdown rendering — headings, bullet/numbered lists, blockquotes, horizontal rules, bold, italic, code, and links all rendered in assistant messages
• Active file context — includes the currently open file and cursor position
• @ references — @file:path, @folder:path, @symbol:name for precise context inclusion
• Image support — paste screenshots or attach images for vision models
• Slash commands — /reset, /undo, /export, /model, /help, /batch, /doc, /spec, /insight, /save, /sessions, /scan, /usage, /context, /test, /lint, /deps, /scaffold, /commit, /verbose, /prompt — with autocomplete dropdown as you type
• Diagram generation — models can generate Mermaid diagrams in code blocks; rendered natively in chat with syntax highlighting and copy-to-SVG support
• Actionable errors — classified error cards with retry, start Ollama, and settings buttons
• Sticky scroll — auto-scroll pauses when you scroll up, floating button to jump back down
• Chat history persistence — conversations survive VS Code restarts (per-workspace)
• Streaming indicator — shows token count and generation speed
• Model management — switch models, install new ones from Ollama

Multi-Backend Support

• Ollama (default) — runs locally, free, no API key needed
• Anthropic API — use Claude models with your API key, with prompt caching for ~90% input token cost reduction
• OpenAI-compatible — works with LM Studio, vLLM, llama.cpp, text-generation-webui, OpenRouter, and any server with a /v1/chat/completions endpoint
• Same interface for all — just change sidecar.baseUrl and optionally sidecar.provider

MCP (Model Context Protocol)

• Connect to any MCP server for external tools (Gmail, Slack, databases, custom tools)
• MCP tools appear transparently alongside built-in tools
• Configure via sidecar.mcpServers setting:

  "sidecar.mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/dir"]
    }
  }
  

Security Scanning

• Automatic secrets detection — scans files after agent writes for AWS keys, GitHub tokens, API keys, private keys, JWTs, connection strings, and more
• Vulnerability scanning — flags SQL injection, command injection, XSS (innerHTML), eval usage, and insecure HTTP URLs
• Integrated into diagnostics — get_diagnostics includes security findings alongside compiler errors
• Pre-commit scan — /scan command or SideCar: Scan Staged Files for Secrets in the command palette scans staged git files before committing
• Skips comments, node_modules, lock files, and minified code

Tool Registry (20+ built-in tools + MCP)

Project Instructions (SIDECAR.md)

Create a SIDECAR.md file in your project root to give SideCar project-specific instructions that persist across sessions:

# Project: My App

## Build
- Run `npm run build` to compile
- Run `npm test` to run tests

## Conventions
- Use TypeScript strict mode
- Prefer async/await over callbacks
- Components go in src/components/

SideCar reads this file on every message and includes it in the system prompt.

Hooks

Run shell commands before/after any tool execution:

"sidecar.hooks": {
  "write_file": { "post": "npm run lint --fix" },
  "*": { "pre": "echo \"Tool: $SIDECAR_TOOL\"" }
}

Environment variables: SIDECAR_TOOL, SIDECAR_INPUT, SIDECAR_OUTPUT (post only).

Scheduled Tasks

Run recurring agent tasks on an interval:

"sidecar.scheduledTasks": [
  { "name": "Lint check", "intervalMinutes": 30, "prompt": "Run the linter and fix any issues", "enabled": true }
]

Scheduled tasks run autonomously and log to the SideCar Agent output channel.

GitHub Integration

• Clone repos, list/view/create PRs and issues
• View commit history, diffs, push/pull
• Browse repo files on GitHub

Requirements

• Ollama installed and in your PATH (for local models)
• Visual Studio Code 1.88.0 or later

Getting Started

1. Install Ollama if you haven't already
2. Install the SideCar extension
3. Click the SideCar icon in the activity bar
4. Start chatting — SideCar launches Ollama automatically

Using with Anthropic API

1. Set sidecar.baseUrl to https://api.anthropic.com
2. Set sidecar.apiKey to your Anthropic API key
3. Set sidecar.model to a Claude model (e.g. claude-sonnet-4-6)

Using with OpenAI-compatible servers

Works with LM Studio, vLLM, llama.cpp, text-generation-webui, OpenRouter, and more:

1. Set sidecar.baseUrl to your server URL (e.g. http://localhost:1234)
2. Set sidecar.apiKey if your server requires it (optional for most local servers)
3. Set sidecar.model to the model name on your server

SideCar auto-detects the provider. To override, set sidecar.provider to "openai".

Keyboard Shortcuts

Extension Settings

Documentation

Full documentation is available at nedonatelli.github.io/sidecar.

Support & Contact

• Bug reports & feature requests: GitHub Issues
• Email: sidecarai.vscode@gmail.com
• Documentation: nedonatelli.github.io/sidecar

Disclaimer

SideCar is an independent project by Nicholas Donatelli and is not affiliated with, endorsed by, or sponsored by Ollama, Anthropic, Meta, Mistral AI, Google, GitHub, or any other company. All product names are trademarks of their respective holders.

License

MIT