SideCar Logo

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 still sets SideCar apart is the combination underneath: genuinely free and fully local-first (runs offline on Ollama, no required cloud), a scaffolding harness that makes weaker local models usable, deep native VS Code integration, and a safety/parallelism stack (shadow workspaces, audit mode, fork & parallel solve, facets) — without a subscription.

vs. open-source agents

SideCar's closest competitors are the other free, open-source agentic extensions: Cline, Kilo Code (the Cline/Roo hybrid most Roo users adopted after Roo Code shut down in May 2026), and Continue (now a full agent, not just autocomplete). SideCar and all three run local models offline, support MCP, and ship a project-rules file — so agent mode itself is no longer a differentiator. The honest differences (verified June 2026):

Where SideCar is different

Inline completions — SideCar has Copilot-style FIM autocomplete built in; Cline and Kilo are agent-only (their users run a separate completion tool). Continue has its own completions too.
Local-model scaffolding harness — capability-adaptive gates and reprompt budgets (Verify / Adapt / Orchestrate / Measure) plus an ablation harness (npm run eval:ablation) to make weaker local models usable. The others assume a capable model; none has an equivalent.
Parallelism & write-isolation infra — shadow workspaces (ephemeral git worktree), audit mode (in-memory write buffer), fork & parallel solve (N attempts, pick-the-winner), typed sub-agent facets, and a local model arena with ELO.
Deep VS Code integration — Problems-panel security/diagnostics, Test Explorer agent runs, CodeLens actions, file decorations, and an activity-bar review badge.

Where they're ahead

Reach — Cline and Kilo run across multiple editors (JetBrains, Zed, Cursor, …) plus CLI/SDK; SideCar is VS Code-only.
Ecosystem & maturity — much larger communities, MCP/plugin marketplaces, and (Cline/Kilo) an optional managed pay-as-you-go provider.

Different form factor: Aider and Goose are strong open-source agents too, but they're terminal/CLI tools (Aider is git-native; Goose is CLI + desktop), not VS Code extensions — a different workflow rather than a head-to-head. Pure-autocomplete extensions like Twinny do single-line completion only; SideCar does that plus the full agent. (Llama Coder, the old Ollama-autocomplete option, appears unmaintained as of 2026.)

vs. proprietary tools

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

Capability	SideCar	Copilot	Cursor	Claude Code
Autonomous agent loop	Yes	Yes	Yes	Yes
Model agnostic (any provider)	Yes	Partial	Partial	No
Fully offline / self-hosted	Yes	No	No	No
HuggingFace model install	Yes	No	No	No
Custom skills system	Yes	Yes	Yes (.cursorrules)	Yes
Context compaction (manual + auto)	Yes	Yes	Yes	Yes
Spending budgets & cost tracking	Yes	Yes	Yes	Yes
Hybrid local-worker delegation	Partial	No	No	No
Prompt pruner & pre-request caching	Yes	No	No	Partial
Plan-then-execute mode	Yes	Yes	Yes	Yes
Review mode (batch diff review)	Yes	No	Partial	No
Native Problems panel integration	Yes	No	No	No
Test Explorer integration (agent test runs)	Yes	No	No	No
Inline diff streaming in chat	Yes	No	No	No
Status bar health indicator	Yes	Partial	No	No
Getting-started walkthrough	Yes	Yes	No	No
Native modal approval for destructive tools	Yes	Partial	Partial	Partial
Conversation steering (type while processing)	Yes	No	Yes	Yes
Works in your existing VS Code	Yes	Yes	No (fork)	Yes (extension + CLI)
MCP support	Yes (client + server)	Yes	Yes	Yes
Monthly subscription	Free	Free tier; $10–39/mo	from $20/mo	Usage-based / from $20/mo

Cursor is a standalone VS Code fork; Windsurf — the other major AI IDE — was rebranded to Devin Desktop by Cognition in 2026. Both are paid editors you switch into; SideCar (like Copilot, Cline, and Continue) runs inside the VS Code you already use.

What sets SideCar apart

Free & fully local-first — runs entirely offline on local Ollama models with no API key, no required cloud, and no subscription. Plug in Anthropic / OpenAI / AWS Bedrock / etc. only when you want a frontier model.
Built for weaker local models — a capability-adaptive scaffolding harness (grounded review + citation gates, tiered reprompt budgets, write/rewrite-thrash defenses) with an ablation harness (npm run eval:ablation) to keep only the scaffolds that earn their latency. This is the differentiator most agents lack — they assume a capable model.
Parallelism & write-isolation — shadow workspaces (ephemeral git worktree), audit mode (atomic in-memory flush), fork & parallel solve (N attempts, pick the winner), typed sub-agent facets, and a local model arena with ELO.
Feels first-party — status bar health indicator, native error toasts, lightbulb code actions, Problems-panel integration, file decorations, activity-bar badge, Test Explorer agent runs, and a SideCar: command palette category.
No vendor lock-in — Ollama, Anthropic, AWS Bedrock, OpenAI-compatible (LM Studio, vLLM, OpenRouter), Kickstand, Groq, Fireworks, Gemini, Copilot, or GGUF directly from HuggingFace.
Hybrid cost-aware — Anthropic prompt caching + 90%-reduction prompt pruner + delegate_task to a local Ollama worker + session spend tracker + daily/weekly budgets + architect/editor model split (sidecar.editorModel) that auto-routes execution turns to a faster/cheaper model.
Security from the ground up — OS keychain key storage, secrets detection, vuln scanning, path-traversal protection, workspace hook warnings, macOS Seatbelt sandbox for agent shell commands.
Agentic across six modes — cautious / autonomous / manual / plan / review / audit, with Copilot-style inline completions and MCP (client and server) on top.
Extensible — MCP (stdio / HTTP / SSE) with lazy tool-schema loading (~50% less context per server) and round-trip write verification, custom skills via markdown, 11 built-in skills, NoSQL quick-install for MongoDB + Redis.

Features

Feature	Description
86 built-in tools	File ops, shell, git, web search, database (SQLite/PG/MySQL/DuckDB), screenshot, doc-to-test synthesis, code profiling, LaTeX compilation, MCP task delegation, research assistant, CI failure analysis, change-impact analysis, numerical-contract checking, and more — full list
Review / Audit modes	Buffer all writes in-memory; review per-file diffs before anything touches disk. Audit adds atomic flush + deletion coverage
Shadow Workspaces	Run tasks in an ephemeral `git worktree` — main tree untouched until you accept
Typed Sub-Agent Facets	Dispatch named specialists (`test-author`, `security-reviewer`, etc.) in parallel, each with its own tool allowlist and preferred model
Fork & Parallel Solve	`/fork <task>` runs N parallel agent approaches; pick-the-winner review UI
NotebookLM research mode	`/notebook` — source-grounded chat with mandatory inline citations and five study-aid generators
Project Knowledge Index	Semantic search over every function/class in your workspace via tree-sitter + MiniLM embeddings; chunk-level retrieval for prose docs
Change-impact analysis	A consequence-aware code graph (AST-exact call/type-flow edges for TS + Python). `analyze_impact` answers "what depends on this symbol?" — import-resolved callers, type-users, subtypes, importers. Opt-in `sidecar.codeGraph.impactGate` blocks finishing with unverified cross-file dependents
Numerical-correctness contracts	For scientific Python: `check_numerical_contracts` finds array kernels (`np.ndarray`/`NDArray`/tensors) lacking a shape/dtype contract; `check_shape_consistency` propagates shape specs (jaxtyping/nptyping/numpy) and flags provable conflicts — intra-kernel, tail-call, and cross-call dataflow. Opt-in `sidecar.numericalContracts.gate` turns "tests pass" into "the array contracts are stated"
Active file context bar	Pill above the chat input — one click includes or excludes the currently open file from agent context
Adversarial Critic	Second LLM call after every edit that finds bugs, regressions, and security issues — blocks the turn on high-severity findings
Completion gate	Blocks the agent from declaring done until lint and colocated tests for edited files have actually run — gates run via the agent's real execution path and key on output evidence, not exit codes
Local-model scaffolding harness	Capability-adaptive machinery that makes weaker local models usable: grounded review + citation-resolution gates, constrained-decoding repair of malformed tool calls, capability-tiered reprompt/critic/compression budgets (weak models get more help, strong models run lean), read-only specialist routing, and write/rewrite-thrash defenses. A failure taxonomy buckets every run (`malformed-call` / `wrong-tool` / `timeout` / `incomplete` / `bad-reasoning`) and `npm run eval:ablation` measures each scaffold's pass-rate lift vs. latency so pure-tax scaffolds get cut
SIDECAR.md	Path-scoped project instructions — sections inject only when the active file matches their `@paths` glob; per-directory files cascade root-to-leaf; falls back to `AGENTS.md` / `CLAUDE.md` / `.cursorrules` when no SIDECAR.md is present
DESIGN.md	Always-injected architecture / style guide — place `.sidecar/DESIGN.md` (or `DESIGN.md`) to keep domain knowledge in every system prompt without SIDECAR.md boilerplate
Inline completions	Copilot-style autocomplete via Ollama FIM or Anthropic (opt-in via `sidecar.enableInlineCompletions`)
Inline chat	`Cmd+I` — edit code in place; lightbulb integration surfaces Fix / Explain / Refactor on diagnostics
Model Arena	`/arena` opens a side-by-side panel comparing 2–4 models on the same prompt with live ELO ratings; `/arena agent` runs a task through different models via fork dispatch
RAM/VRAM monitor	Live status bar item shows free RAM % and GPU % (colour-coded); warns or blocks before Kickstand model loads, HuggingFace installs, and arena sessions when memory is critically low. Click to refresh, or `sidecar.memory.refresh` command
Selective regeneration	Select any text in an assistant response → a bar appears to rewrite just that section with optional instruction, in-place
Background agents	`/bg <task>` spawns parallel autonomous agents with a status dashboard; toast notification + status bar spinner on completion
External context providers	Pull live GitHub Issues, Linear, Jira, or Bitbucket Cloud tickets into every agent system prompt via `sidecar.contextProviders`
CI Failure Analysis	`analyze_ci_failure` agent tool + `/ci` slash command parses failing GitHub Actions logs, windows the relevant error region, and proposes a fix. Gated by `sidecar.ciAnalysis.enabled`
Regression Guards	`/guards` slash command + `RegressionGuardHook` runs ecosystem-aware post-edit checks (`lint-clean`, `tests-pass`, `no-new-todos`). Declare per-skill guards via `guards:` frontmatter
Monorepo Support	`monorepo_packages` tool auto-detects Nx / Turbo / pnpm workspaces / Yarn / Lerna and lists workspace packages with metadata. Gated by `sidecar.monorepo.enabled`
macOS Seatbelt sandbox	`sidecar.sandbox.enabled` wraps agent shell commands with a deny-default SBPL profile — writes restricted to the workspace + `/tmp` + build caches
Dependency Drift Alerts	Scans `package.json`, `requirements.txt`, `Cargo.toml`, and `go.mod` for outdated and vulnerable deps — CVE/GHSA IDs from the OSV API surface in the Problems panel; `check_dependencies` agent tool for on-demand reports
Test Current Model	`sidecar.testCurrentModel` runs the smoke eval suite against the active model, shows a progress notification, and reports pass/fail counts — Open Report button opens structured per-case failure details
Security scanning	Secrets, SQL injection, XSS, eval — findings in VS Code Problems panel (`source:sidecar-*`)
Doc sync	Checks JSDoc `@param` staleness and README call-site arity on every save
MCP	stdio / HTTP / SSE transports, per-server tool allowlist, `sidecar.noSql.install` for one-click MongoDB + Redis
Skills 2.0	Constrained skills: `allowed-tools`, `preferred-model`, `max-iterations`, and `disable-model-invocation` frontmatter fields enforce execution contracts per skill; 🛡 badge in QuickPick for restricted skills
Skills Registry & Sync	`/skills` opens a searchable QuickPick with registry-origin tags, tool-allowlist chips, and Stack mode for composing skills. `SideCar: Publish Skill to Registry` commits and pushes a skill to your git registry. `SideCar: Sync Skill Registries` pulls without restart. `hourly`/`daily` autoPull for background sync
Chat Threads & Branching	`/branch [name]` forks the current conversation — original preserved, branch continues independently; nested parent/child tree in the Sessions sidebar
Research Assistant	Structured project tracking: 8 agent tools (`research_create_project`, `_add_hypothesis`, `_log_experiment`, `_add_observation`, `_update_hypothesis_status`, `_set_project_status`, `_list_projects`, `_export_report`) + sidebar panel + `/research` slash command. Gated by `sidecar.research.enabled`
Context window bar	3 px colour-coded fill bar above the input area — blue → yellow (≥ 60 %) → red (≥ 80 %). Tooltip shows `Context: 12K / 32K tokens (38%)`
Message editing	Click ✎ on any user message to edit and resend — subsequent messages fade out as a truncation preview, then history is rewound and the agent re-runs with the edited text
Skills	11 built-in skills (review-code, debug, refactor, write-tests, break-this, explain-code, create-skill, mcp-builder, tdd-red-green-refactor, typed-service-contracts, agent-dx-cli-scale) + custom markdown skills
Slash commands	`/model`, `/fork`, `/branch`, `/compact`, `/notebook`, `/bg`, `/pr`, `/review`, `/ci`, `/memories`, `/commit`, `/doc`, `/spec`, `/revise`, and 25+ more with autocomplete

Requirements

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

Getting Started

Ollama (local, free)

Install Ollama if you haven't already
Install the SideCar extension
Click the SideCar icon in the activity bar
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, highest agent eval at 98%). Low-RAM: granite4.1:3b (2 GB, 81%). Cloud: configure the Anthropic backend for maximum reliability.

Anthropic (Claude)

Click the ☰ menu in the chat header → Anthropic Claude under Backend
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

Set sidecar.baseUrl to https://api.openai.com
Run SideCar: Set / Refresh API Key and paste your OpenAI key
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.

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

OpenRouter

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

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

Google Gemini

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

Groq

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

Set sidecar.baseUrl to https://api.groq.com/openai
Run SideCar: Set / Refresh API Key and paste your Groq key
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.

Run Kickstand locally — it auto-generates a bearer token at ~/.config/kickstand/token
Set sidecar.baseUrl to your Kickstand URL (default http://localhost:4000)
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:

Set sidecar.baseUrl to your server URL (e.g. http://localhost:1234)
Set sidecar.model to the model name on your server
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.

Model	Size	Notes
gemma4:e4b (default)	9.6 GB	Strongest prompt-following of the local models; the shipped default
ministral-3:latest	6.0 GB	Lightest strong agent; recommended low-footprint default
qwen2.5-coder:7b	4.7 GB	Reliable coding baseline — the most consistent tool-caller of the 7B tier
qwen3-coder:30b	18 GB	Best larger local coder (MoE, ~3.3B active); recommended if it fits
qwen3.5:latest	6.6 GB	Reliable as of v0.116 (earlier versions hit Ollama tool-parser issues)
deepseek-r1:8b / qwen3:8b	5.2 GB	Reasoning models; answers surfaced correctly as of v0.116
granite4.1:3b	2.1 GB	Low-RAM option; punches above its weight
llama3.2:latest	2.0 GB	Low-RAM general model
devstral:24b	14 GB	SWE-tuned; supported as of v0.116 (`{tool, args}` text-call parsing fix)

Memory guidance (36 GB). Sub-10 GB models run comfortably alongside VS Code with headroom for the context cache. A 14B model at a large context window (e.g. qwen2.5-coder:14b with SideCar's full num_ctx) can exceed 36 GB of unified memory and thrash — lower sidecar.ollama.numCtx, or prefer the 7B coder. On more memory, the 24–30 GB-class models above have proportionally more room.

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

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

Extension Settings

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

Setting	Default	Description
`sidecar.baseUrl`	`http://localhost:11434`	API base URL
`sidecar.model`	`gemma4:e4b`	Model for chat
`sidecar.agentMode`	`cautious`	`cautious` / `autonomous` / `manual` / `plan` / `review` / `audit` or a custom mode name
`sidecar.agentMaxIterations`	`50`	Max agent loop iterations
`sidecar.agentMaxTokens`	`200000`	Max tokens per agent run
`sidecar.mcpServers`	`{}`	MCP server definitions (or use `.mcp.json`)
`sidecar.toolPermissions`	`{}`	Per-tool overrides: `allow` / `deny` / `ask`
`sidecar.hooks`	`{}`	Pre/post execution hooks per tool
`sidecar.completionGate.enabled`	`true`	Block agent finish until lint + tests pass
`sidecar.enableInlineCompletions`	`false`	Copilot-style autocomplete (opt-in)
`sidecar.mcpDelegation.enabled`	`false`	Enable `delegate_to_mcp` tool (agent → MCP server)
`sidecar.mcpDelegation.allowedServers`	`[]`	Allowlist of servers `delegate_to_mcp` may target (empty = all)
`sidecar.mcpServer.enabled`	`false`	Expose SideCar's agent loop as a local MCP server
`sidecar.mcpServer.port`	`3457`	Port for the SideCar MCP server (127.0.0.1 only)
`sidecar.mcpServer.requireAuth`	`false`	Require bearer token for inbound MCP requests
`sidecar.dailyBudget`	—	Daily spend cap in USD (paid backends)
`sidecar.weeklyBudget`	—	Weekly spend cap in USD (paid backends)

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: 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