SideCar — AI Coding Assistant for VS Code

SideCar is a free, open-source coding agent for VS Code: a full agent loop with tool use, inline completions and inline chat, not just a chat panel. Run it on local Ollama models with no API key and no data leaving your machine, or connect Anthropic, OpenAI, OpenRouter, Groq, Fireworks, Google Gemini, AWS Bedrock, your GitHub Copilot subscription, or any OpenAI-compatible server.

Quick start

1. Install Ollama and pull the default model: ollama pull gemma4:e4b
2. Install the SideCar extension
3. Click the SideCar icon in the activity bar (or Cmd+Shift+I / Ctrl+Shift+I)
4. Start chatting. SideCar starts Ollama if it isn't running, and offers Install Model if the selected model is missing.

gemma4:e4b (9.6 GB, about 10 GB of VRAM) is the most-tested local model. On smaller GPUs, try ministral-3:latest (6 GB) or granite4.1:3b (2 GB).

Other providers

Open the ☰ menu in the chat header → Backend (or run SideCar: Switch Backend) and pick a provider. SideCar sets the URL, provider and a default model, then asks for the API key and keeps it in the OS keychain, one slot per provider. Then pick the model you want with SideCar: Select Model.

Provider	What you need
Anthropic, OpenAI, Google Gemini	An API key from the provider (billed separately from Claude.ai or ChatGPT plans)
OpenRouter, Fireworks	An API key; one key reaches many models
Groq	An API key on a paid tier: one SideCar request is larger than the free tier's tokens-per-minute cap
AWS Bedrock	Your AWS credentials or a Bedrock API key, and model access enabled in the region
GitHub Copilot	The Copilot extension, signed in; no separate key
Custom endpoint	Any OpenAI-compatible server (vLLM, LM Studio, llama.cpp, a gateway) and its URL

Setup, tuning and limits for each: docs/backends.md.

Features

The highlights. Full list: feature specs.

Feature	Description
87 built-in tools	File ops, shell, git, web search, databases, screenshots, code profiling, CI-failure analysis, change-impact analysis, and more — full list
Local-model scaffolding harness	Capability-adaptive machinery for weaker local models: tool-call recovery from model text, grounding/completion gates, tiered reprompt budgets, and a trimmed tool catalog for local backends. Net effect is measured, not assumed
Completion gate	Holds back "done" until lint and colocated tests have actually run, reading their output rather than trusting exit codes; failing tests get the agent sent back to fix them, and edited files must parse
Write isolation	Review/Audit modes (buffer writes in-memory, approve per-file diffs) and Shadow Workspaces (run in an ephemeral git worktree; main tree untouched until you accept)
Safety checks	Asks before commits that hold ignored or oversized files, before pushes (naming the remote and whether it's public), and before config changes you didn't ask for
Parallelism	Fork & parallel solve (N attempts, pick the winner), typed sub-agent facets (named specialists with their own tool allowlist + model), and background agents
Durable instruction memory (A/B-proven)	Standing instructions survive context compaction and the session itself, applied automatically in later sessions; review or forget them in the Remembered Instructions view
Project knowledge & impact	Semantic search over every symbol (tree-sitter + embeddings) and an AST-exact code graph — analyze_impact answers "what depends on this symbol?"
Inline completions & chat	Copilot-style FIM autocomplete and Cmd+I in-place editing with Fix/Explain/Refactor on diagnostics
MCP (client + server)	stdio / HTTP / SSE transports with lazy schema loading; expose SideCar's own agent loop as an MCP server

Why SideCar

• Free and local-first. Runs offline on Ollama with no account or key; add a cloud provider only when you want one. No subscription.
• Works with local models that other agents can't drive. Many capable local models write their tool calls as text or JSON in the reply instead of the structured field, so a stock agent sees no call and stalls. SideCar parses and runs them; qwen2.5-coder, for one, can't call tools at all without it.
• Any provider, in the VS Code you already use. Not a fork; switching provider is one menu pick, and you can install GGUF models from Hugging Face.

The trade-off: SideCar is VS Code-only, while Cline and Kilo Code cover more editors and have larger plugin ecosystems.

Measured, not claimed

SideCar carries a lot of machinery meant to make weaker local models usable: gates, reprompt budgets, tool-call repair. Specific repairs are proven to matter (the text-call parsing above, for example). That the stack as a whole makes a local model better at real tasks is not established, and this README doesn't claim it. Mechanisms that measure net-negative are removed. The main gates can be switched off (sidecar.completionGate.enabled, sidecar.syntaxGate.enabled, sidecar.redCheckGate.enabled).

Tested models

Each model below runs SideCar's agent smoke suite (read, edit, write, run, plan and error-recovery tasks) with zero infrastructure errors: no crashes, dropped tool calls or retry loops. How well each completes the tasks varies, and the notes reflect both.

Model	Size	Notes
gemma4:e4b (default)	9.6 GB	Strongest prompt-following of the local models; the shipped default
gemma4:31b	20 GB	The most capable local model tested; needs a 32 GB GPU
ornith:9b	5.6 GB	Best small-footprint agent — near-default capability at a third the size
ministral-3:latest	6.0 GB	Strong, low-footprint, reliable
qwen2.5-coder:7b	4.7 GB	Solid coding baseline; emits text-form tool calls (SideCar's parser recovers them)
north-mini-code:1.0	19 GB	Agentic-tuned coder; strong autonomous, but rarely asks clarifying questions
laguna-xs-2.1	20 GB	Fast MoE coder; some run-to-run variance
granite4.1:3b	2.1 GB	Low-RAM option; punches above its weight
llama3.2:latest	2.0 GB	Low-RAM general model; below the agent floor for complex tasks

Not recommended: lfm2.5, which in testing reliably followed instructions planted in a code block (a prompt injection).

Memory and GPU guidance, and per-model detail: Model Compatibility. In the cloud, Anthropic Claude and Google Gemini are the strongest tested options.

Keyboard shortcuts

Shortcut	Action
Cmd+Shift+I / Ctrl+Shift+I	Toggle SideCar chat panel
Cmd+I / Ctrl+I	Inline chat (edit code in place)
Cmd+L / Ctrl+L	Clear chat
Cmd+Shift+U / Ctrl+Shift+U	Undo all AI changes
Cmd+Shift+E / Ctrl+Shift+E	Export chat as Markdown

Settings

The ones most people change. Every setting: configuration reference.

Setting	Default	Description
sidecar.model	gemma4:e4b	Model for chat and agent runs
sidecar.baseUrl	http://localhost:11434	API base URL
sidecar.agentMode	cautious	cautious / autonomous / manual / plan / review / audit, or a custom mode
sidecar.completionGate.enabled	true	Hold back "done" until lint and tests pass
sidecar.enableInlineCompletions	false	Copilot-style autocomplete (opt-in)
sidecar.dailyBudget	0 (off)	Daily spend cap in USD on paid backends (sidecar.weeklyBudget for weekly)

API keys live in VS Code SecretStorage (the OS keychain), set with SideCar: Set / Refresh API Key, never in settings.json.

Agents Window (preview): to use SideCar in VS Code's Agents Window, install it in your default profile and add to settings.json:

"extensions.supportAgentsWindow": {
    "nedonatelli.sidecar-ai": true
}

Documentation

Full documentation: nedonatelli.github.io/sidecar

• docs/backends.md — setup, tuning, and limits for every backend
• docs/model-compatibility.md — which local models work for agent mode, hardware guidance, and which to avoid
• SECURITY.md — threat model, secret-pattern catalog, vulnerability disclosure
• docs/extending-sidecar.md — skills, custom tools, MCP servers, policy hooks
• CHANGELOG.md — per-release notes
• CONTRIBUTING.md — building, testing, and contributing
• Architecture diagrams: agent loop · tool dispatch · context pipeline · MCP lifecycle

Support

• Bugs and feature requests: GitHub Issues
• Security issues: private disclosure via SECURITY.md; please don't open public issues for vulnerabilities
• Email: [email protected]

SideCar is free and always will be. If it saves you time, a tip is appreciated but never required:

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