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, OpenAI, Fireworks AI, OpenRouter, Google Gemini, Groq, Kickstand (self-hosted), or any OpenAI-compatible server (LM Studio, vLLM, llama.cpp) for AI-powered coding — with full agentic capabilities, inline completions, and tool use.

A free, open-source, local-first autonomous AI agent for coding — a full agent loop, not just chat. No subscription, and with local models, no data leaves your machine.

Why SideCar?

Agent mode, tool use, and MCP are table stakes in 2026 — Copilot, Cursor, Claude Code, Cline, and Continue all have them. What sets SideCar apart is the combination underneath, all with no subscription:

• Genuinely free and fully local-first — runs offline on Ollama with no API key or required cloud; plug in a frontier provider only when you want one.
• Tool-call recovery from model text — many capable local models emit their tool calls as plain text or JSON in the content field instead of the structured tool_calls field, so a stock harness sees no call and stalls. SideCar parses those out and runs them: qwen2.5-coder scores 0% on the Berkeley Function-Calling AST subset without this layer and ~78% with it. Parameter-name recovery (file→path), foreign tool-name aliases (create_file→write_file), and bare-JSON parsing sit alongside it. See Measured, not claimed.
• Capability-adaptive gates — completion, grounding, and citation gates plus tiered reprompt budgets aimed at making weaker local models usable. Whether the net effect is a lift is measured, not assumed.
• Parallelism & write-isolation — shadow workspaces (ephemeral git worktree), audit mode (in-memory write buffer), fork & parallel solve (N attempts, pick the winner), and typed sub-agent facets.
• First-party VS Code integration — Problems-panel security/diagnostics, Test Explorer agent runs, CodeLens actions, file decorations, inline completions, and inline chat.

The trade-off: SideCar is VS Code-only, where Cline and Kilo Code reach more editors and ship larger plugin ecosystems.

vs. proprietary tools

Copilot, Cursor, and Claude Code are all capable paid agents. The comparison is no longer "agentic or not" — it's local-first, IDE-native, and free vs. cloud-tied and subscription-based.

Cursor and Windsurf (now Devin Desktop) are paid editors you switch into; SideCar, like Copilot and Cline, runs inside the VS Code you already use.

Beyond the differentiators above, SideCar is cost-aware (prompt caching, a 90%-reduction prompt pruner, delegate_task to a local worker, spend tracking, and an architect/editor model split), secure by default (OS-keychain key storage, secrets/vuln scanning, path-traversal protection), and extensible (MCP client + server with lazy schema loading, markdown skills, six agent modes). No vendor lock-in: Ollama, Anthropic, OpenAI-compatible, Kickstand, Groq, Fireworks, Gemini, or GGUF from HuggingFace.

Measured, not claimed

SideCar ships a lot of machinery aimed at making weaker local models usable —
gates, reprompt budgets, deterministic tool-call repair, recovery paths. How
much of it actually helps is not yet established, and this section exists so
the README does not imply otherwise.

What is measured today:

• Tool-surface expressibility (tests/llm-eval/toolSurface.eval.ts) — one
  model call per case against the real advertised schemas. Across nine local
  models, seven score 7/7 on picking the right tool and expressing the intent.
  Two repairs are confirmed load-bearing: stringified-array coercion for
  ask_user (llama3.2) and text-protocol call parsing (qwen2.5-coder emits
  calls as JSON in the content field, and would score 0/7 without it).
• Individual mechanisms with recorded A/Bs — noted inline where they exist,
  with the model and date, because a result on one model is not a general claim.

What is not measured:

• The net effect of the scaffolding stack as a whole. No end-to-end
  scaffold-on vs scaffold-off comparison has been run on current code.
• Per-mechanism cost. Reprompts and recovery paths consume context and latency;
  which ones earn that is an open question, and at least one recovery path
  (splitFusedAnchor) was measured doing active harm and was removed in
  scaffold 4.0.0.
• Anything about how these mechanisms behave on models or tasks outside the
  small local set that has been dogfooded.

Historical eval percentages have been removed from this README rather than
restated: every local number recorded before 2026-07-30 was measured with
thinking disabled — a configuration that is not the shipped default — so those
figures do not describe what a user gets. Re-baselining is in progress.

Features

The highlights — full feature list in the docs.

Requirements

• Visual Studio Code 1.116.0 or later
• Ollama installed and in your PATH (for local models only)

Getting Started

Ollama (local, free)

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 and downloads a starter model

Default model: gemma4:e4b (9 GB, ~10 GB VRAM — the most-dogfooded local model, with the strongest prompt-following of those tested and a cold-start the harness handles automatically). Lighter alternative: ministral-3:latest (6 GB, 8 GB VRAM). Low-RAM: granite4.1:3b (2 GB). Cloud: configure the Anthropic backend for maximum reliability.

Anthropic (Claude)

1. Click the ☰ menu in the chat header → Anthropic Claude under Backend
2. SideCar prompts for your API key on first switch — paste it and you're done

Or manually: sidecar.baseUrl = https://api.anthropic.com, run SideCar: Set / Refresh API Key, set sidecar.model = claude-sonnet-4-6.

SideCar uses Anthropic's prompt caching automatically — subsequent turns in a session cost ~90% less on input tokens.

OpenAI

1. Set sidecar.baseUrl to https://api.openai.com
2. Run SideCar: Set / Refresh API Key and paste your OpenAI key
3. Set sidecar.model to gpt-5 (recommended — higher TPM than mini models)

Note: gpt-4o-mini and gpt-4.1-mini share a 200K TPM org-level cap. At ~23K tokens per request (system prompt + tools), the budget exhausts after ~8 requests. Use gpt-5 or a model with a higher TPM allocation.

Fireworks AI

Fireworks offers fast inference for large open-source models including DeepSeek and Llama variants.

1. Set sidecar.baseUrl to https://api.fireworks.ai/inference
2. Run SideCar: Set / Refresh API Key and paste your Fireworks key
3. Set sidecar.model to accounts/fireworks/models/deepseek-v4-pro (recommended)

OpenRouter

OpenRouter provides a single API key for 100+ models across providers.

1. Set sidecar.baseUrl to https://openrouter.ai/api
2. Run SideCar: Set / Refresh API Key and paste your OpenRouter key
3. Set sidecar.model to e.g. x-ai/grok-3-mini or google/gemini-2.5-flash

Google Gemini

1. Set sidecar.baseUrl to https://generativelanguage.googleapis.com/v1beta/openai
2. Run SideCar: Set / Refresh API Key and paste your Gemini API key
3. Set sidecar.model to gemini-2.5-flash

Groq

Groq offers very fast inference for open-source models via their LPU hardware.

1. Set sidecar.baseUrl to https://api.groq.com/openai
2. Run SideCar: Set / Refresh API Key and paste your Groq key
3. Set sidecar.model to e.g. llama-3.3-70b-versatile

Note: Groq's free tier (12K–30K TPM) is exhausted after 1–2 requests from SideCar's system prompt + tool schemas. A paid tier is required for reliable use.

Kickstand (self-hosted manager)

Kickstand is a self-hosted model manager that wraps Ollama with a management API, model registry, and load/unload controls.

1. Run Kickstand locally — it auto-generates a bearer token at ~/.config/kickstand/token
2. Set sidecar.baseUrl to your Kickstand URL (default http://localhost:11435)
3. SideCar reads the token file automatically — no API key prompt needed

Other OpenAI-compatible servers

LM Studio, vLLM, llama.cpp, and other OpenAI-compatible servers work out of the box:

1. Set sidecar.baseUrl to your server URL (e.g. http://localhost:1234)
2. Set sidecar.model to the model name on your server
3. Run SideCar: Set / Refresh API Key if your server requires authentication

SideCar auto-detects the provider from the URL. Override with sidecar.provider: "openai" if needed.

Tested Models

SideCar is verified against an agent smoke-eval suite (read / edit / write / run / plan / error-recovery tasks) on each supported model. The primary gate is infrastructure reliability — the agent loop, tool-call parsing, and completion gates must run cleanly (no crashes, dropped tool calls, or retry thrash) — measured separately from a model's raw task capability.

Test hardware: Apple M3 Max, 36 GB unified memory (macOS), Ollama for local models. Model sizes and memory guidance below assume this configuration.

Local models (Ollama)

Every model below runs the agent smoke suite with zero infrastructure errors. Task-completion capability varies; the recommendations reflect both.

Memory guidance (36 GB). Sub-10 GB models run comfortably alongside VS Code with headroom for the context cache; the ~19–20 GB MoE coders fit with less room to spare. A dense 14B model at a large context window can exceed unified memory and thrash — lower sidecar.ollama.numCtx, or prefer a smaller coder.

Cloud models

For maximum reliability, the Anthropic (Claude), Google Gemini, and Fireworks (DeepSeek) backends are the strongest tested cloud options; see the setup sections above.

VS Code Copilot Chat & Agents Window

SideCar registers as a native VS Code chat participant — type @sidecar in the Copilot Chat panel to talk to your configured backend without opening the SideCar sidebar. Slash commands /review, /fix, /explain, and /commit-message are available.

VS Code Agents Window (Preview): To use SideCar in the dedicated Agents Window, add the following to your VS Code settings.json:

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

The extension must be installed in your default VS Code profile.

Keyboard Shortcuts

Extension Settings

Core settings — full reference at nedonatelli.github.io/sidecar/configuration.

API keys are stored in VS Code SecretStorage (OS keychain) — set via SideCar: Set / Refresh API Key, never in plaintext settings.

Documentation

Full documentation: nedonatelli.github.io/sidecar

• 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
• Architecture diagrams: agent loop · tool dispatch · context pipeline · MCP lifecycle

Support & Contact

• Bug reports & feature requests: GitHub Issues
• Security issues: private disclosure via SECURITY.md — please don't open public issues for vulnerabilities
• Email: [email protected]
• Documentation: nedonatelli.github.io/sidecar

SideCar is free and always will be. If it saves you time and you'd like to support development, 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