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, AWS Bedrock, your GitHub Copilot subscription, Kickstand (self-hosted), or any OpenAI-compatible server (vLLM, LM Studio, llama.cpp, a gateway) 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, for one, emits its calls as JSON in the content field). 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, syntax, red-check, 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 (Anthropic prompt caching, a prompt pruner that trims oversized tool results, delegate_task to a local worker, spend tracking and budgets, 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 plus custom ones). No vendor lock-in: Ollama, Anthropic, OpenAI, OpenRouter, Groq, Fireworks, Gemini, AWS Bedrock, GitHub Copilot, Kickstand, any OpenAI-compatible endpoint, 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. That the stack as a whole makes a local model better at real tasks is not established, and this README does not claim it.

What has been shown:

• Specific repairs are load-bearing. Text-protocol call parsing is what lets qwen2.5-coder call tools at all (it emits calls as JSON in the content field), and stringified-array coercion is what makes llama3.2's ask_user calls valid.
• Individual mechanisms with recorded A/Bs — noted inline where they exist, because a result on one model is not a general claim.

Mechanisms that cost more than they return are removed rather than kept: the adversarial critic was removed in v0.124.0 after measuring net-negative. Each completion gate has its own on/off setting (sidecar.completionGate.enabled, sidecar.syntaxGate.enabled, sidecar.redCheckGate.enabled, sidecar.behavioralVerificationGate.enabled) if you want to run leaner.

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

Switching backends: open the ☰ menu in the chat header → Backend, or run SideCar: Switch Backend from the Command Palette. Each entry sets sidecar.baseUrl, sidecar.provider, and sidecar.model in one step and keeps its API key in its own SecretStorage slot, so switching never clobbers a key you already set. The per-backend sections below also give the manual settings. Full details for every backend: docs/backends.md.

Ollama (local, free)

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 launches Ollama automatically if it isn't running. If the selected model isn't installed, SideCar offers an Install Model action.

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. Open the ☰ menu in the chat header → Anthropic Claude under Backend
2. With no key stored yet, SideCar offers Set API Key — paste your key from platform.claude.com (or run SideCar: Set / Refresh API Key any time)

The menu entry selects claude-haiku-4-5 (lowest cost); for agent work, claude-sonnet-4-6 is the recommended model. Manually: sidecar.baseUrl = https://api.anthropic.com, then set sidecar.model.

SideCar uses Anthropic's prompt caching automatically — cache hits cost ~90% less on input tokens. The Anthropic API is billed separately from Claude.ai Pro/Max subscriptions.

OpenAI, OpenRouter, Groq, Fireworks, Google Gemini

Pick the provider in the backend menu, then paste its key when SideCar asks (or run SideCar: Set / Refresh API Key). To configure by hand, set sidecar.baseUrl and sidecar.model:

Rate limits: SideCar's system prompt + tool schemas are ~23K tokens per request. gpt-4o-mini and gpt-4.1-mini share a 200K TPM org-level cap that exhausts after ~8 requests — use gpt-5 or a model with a higher TPM allocation. Groq's free tier (12K–30K TPM) is exhausted after 1–2 requests; a paid tier is required for reliable use.

AWS Bedrock

Claude models through Bedrock, with the same tool use, thinking, and streaming as the direct Anthropic backend.

1. Open the ☰ menu → AWS Bedrock under Backend; SideCar asks for the region (change it later with SideCar: Bedrock: Set Region). GovCloud regions are supported.
2. Credentials come from the standard AWS chain — AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY (+ AWS_SESSION_TOKEN) or ~/.aws/credentials — or a Bedrock API key via SideCar: Set / Refresh API Key or AWS_BEARER_TOKEN_BEDROCK.
3. The model list is read live from your account; the model must be enabled for the region in the Bedrock console.

Prompt caching is not sent on the Bedrock path yet. See docs/backends.md.

GitHub Copilot

Route requests through your existing Copilot subscription via VS Code's Language Model API — no separate key. Requires the GitHub Copilot extension, signed in. Pick GitHub Copilot in the backend menu, or set sidecar.provider to copilot. Tool calling depends on the Copilot model; SideCar falls back to chat-only when it isn't supported.

Kickstand (self-hosted)

Kickstand is a self-hosted inference server (CLI: kick) that loads, unloads, and runs GGUF and MLX models with GPU acceleration.

1. Start it with kick start — it auto-generates a bearer token at ~/.config/kickstand/token
2. Pick Kickstand in the backend menu, or set sidecar.baseUrl to http://localhost:11435
3. SideCar reads the token file automatically — no API key prompt needed

Custom endpoint (any OpenAI-compatible server)

vLLM, LM Studio, llama.cpp, mlx-lm, or a gateway: pick Custom endpoint in the backend menu and enter the server's URL, then its API key (leave it empty if the server doesn't need one — no Authorization header is sent). Or configure it by hand:

1. Set sidecar.provider to openai-compat and sidecar.baseUrl to your server URL (e.g. http://localhost:8000; a trailing /v1 is optional)
2. Set sidecar.model to the model name on your server
3. Run SideCar: Set / Refresh API Key if your server requires authentication

SideCar reads the model's context window from the server's /v1/models (max_model_len on vLLM), so history is compacted before a request can overflow it. A server on localhost gets the same prompt-size cap and trimmed tool catalog as local Ollama. vLLM needs --enable-auto-tool-choice and --tool-call-parser for the agent to call tools — see the launch example.

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: local models were first qualified on an Apple M3 Max laptop with 36 GB of unified memory (macOS); current testing runs on a Windows workstation with an NVIDIA RTX 5090 (32 GB VRAM) and 128 GB of system RAM. Both use Ollama for local models. Memory guidance for each is below.

Local models (Ollama)

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

Not recommended: lfm2.5 — in qualification it reproducibly obeyed a fenced prompt injection.

Memory guidance — Apple Silicon, 36 GB unified memory. 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.

Memory guidance — discrete GPU (tested on 32 GB VRAM). What matters is VRAM, not system RAM: the model's weights and its context cache must fit on the GPU, or Ollama moves layers to the CPU and generation slows sharply. The context cache grows with the window — gemma4:e4b at its full 131K context occupies about 9.7 GB of VRAM. If a larger model runs slowly, check ollama ps (the PROCESSOR column should read 100% GPU) and lower sidecar.ollama.numCtx before switching models.

More per-model detail: Model Compatibility.

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. Agent requests follow your sidecar.agentMode and ask before acting in a dialog (Copilot's own tool-approval settings don't apply, because SideCar runs its own tools).

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

• docs/backends.md — setup, tuning, and limits for every backend
• docs/model-compatibility.md — which local models work for agent mode, 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 & Contact

• Bug reports & feature requests: GitHub Issues
• Security issues: private disclosure via SECURITY.md — please don't open public issues for vulnerabilities
• Email: sidecarai.vscode@gmail.com
• 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