Why LocalPilot?
It costs nothing per month, and you can check for yourself that it is private.
| Per developer | LocalPilot | GitHub Copilot Pro / Pro+ / Max | Cursor Plus / Premium / Super | Claude Pro / Max |
|---|
| Per month | $0 | $10 / $39 / $100 | $8 / $30 / $100 | $20 / $100–200 |
| Team of 10, per year | $0 | $1,200 / $4,680 / $12,000 | $960 / $3,600 / $12,000 | $2,400 / $12,000+ |
| Works with the network unplugged | Yes | No | No | No |
| Where your code is processed | Your laptop | Vendor cloud | Vendor cloud | Vendor cloud |
Published list prices for individual plans (USD) at the time of writing; vendors change plans and limits often. With LocalPilot you pay only for your own electricity — a 7B coding model needs ~8 GB RAM, 14B–32B models 16–32 GB.
Three checks, under a minute. Turn off Wi‑Fi: LocalPilot keeps editing and running tests. Open Settings → Providers: every address is localhost unless you added one. Read the source: it is MIT-licensed, and npm test runs against your own Ollama.
- Private and offline. Your code only reaches the model servers you list, and the defaults are
localhost. No account, no telemetry SDK, no update check; the Dashboard's metrics live in ~/.localpilot/telemetry and never leave the disk (localpilot.telemetry.enabled: false turns them off). The only other outbound calls are ones you control: the skill registry index from GitHub (localpilot.skills.autoInstall: off stops it), the web tools (localpilot.approvals.network: ask confirms each request, never blocks them), and model downloads you start yourself.
- A full agent. It reads, searches, edits, runs commands and tests, retries when something fails and asks before anything risky.
- Written for local models. A text fallback for models without tool calling, and context trimming so 8k to 32k windows still finish long tasks.
- Parallel agents. Run several at once, see every file each one touched, keep or undo per agent.
- Pause, steer, stop. Pause any agent after its current step, type new instructions it picks up on resume, or stop it — for the whole run or a single parallel subagent.
- Multitask like Cursor — locally. Split work across agents or run one task on several models and keep the best; every agent works in its own git worktree so nothing collides, and you apply what you like. The launcher suggests tasks for your project and autocompletes as you type.
- Works with your existing setup. Reads
AGENTS.md, CLAUDE.md, .cursorrules, .github/copilot-instructions.md, .claude/skills and .claude/agents as they are.
- Free, MIT licensed.
Install
From the VS Code Marketplace (recommended) — search for LocalPilot Agent in the Extensions view, or:
code --install-extension localpilot.localpilot-agent
From a GitHub release
curl -L -o localpilot.vsix https://github.com/isathish/localpilot/releases/latest/download/localpilot.vsix
code --install-extension localpilot.vsix
Requires VS Code 1.104+.
Serve a model — or let LocalPilot do it: with no model found it offers Set up automatically, which starts or installs Ollama and downloads a starter coder model sized for your machine. Any of these works too (running servers are auto-detected, including ones in Docker):
ollama pull gemma4:e4b && ollama serve # Ollama → :11434
# LM Studio → Developer tab → Start server → :1234
# llama-server -m model.gguf --port 8080 → :8080
# vllm serve google/gemma-4-12B-it → :8000
# docker model pull ai/gemma4:e4b → Docker Model Runner :12434
# ./start-here.sh (github.com/JustVugg/colibri) → colibri :8000, huge MoE models from disk
# docker run -p 11434:11434 ollama/ollama → any model-server container
Open the LocalPilot icon in the Activity Bar (or ⌘⇧L) and ask for a change — Auto picks the best tool-capable model it found.
Features
| |
|---|
| Agent · Ask · Plan modes | Autonomous edit/run loop with approval gates; read-only Q&A; investigate-then-plan with a live Tasks checklist. |
| Parallel agents & multitasking | Chat tabs run independently; Run Tasks in Parallel (⌘⇧⌥P) spawns one agent per task. The Agents panel shows live status and changed files. |
| Pause / steer / stop | Pause (⌥Esc) holds an agent after its current step; anything you type while it is paused or running is queued and read at its next step; Stop (Esc) cancels. Each parallel subagent lane has its own pause and stop; pausing or stopping the parent reaches all of them. Also in the tab menu, background strip, Agents panel, Command Palette and the CLI (Ctrl+Z / Ctrl+C). |
| Change review & checkpoints | Every agent edit is snapshotted — Diff / Accept / Reject per file or per session, and Restore checkpoint on any message rewinds files and conversation. |
| Zero-config providers | Auto-detects Ollama, LM Studio, llama.cpp, vLLM, colibri, Docker Model Runner, Jan, SGLang, KoboldCpp and model servers running in Docker containers; new ones are added to your settings automatically. |
| Huge MoE models with colibri | A running coli serve is recognised even on vLLM's port :8000. LocalPilot reads its capabilities (vision, chat vs decision-only models) and KV slots, and judges memory fit by the part that stays in RAM, not the 20–1,500 GB download, so GLM-5.2, DeepSeek V4 Flash or Qwen3-Coder-30B route like any other model. Models whose experts stream from disk get patient stall timeouts and are kept for Plan mode rather than chatty subagents or ghost text. LocalPilot: Start colibri launches the server from your checkout (localpilot.colibri.*). |
| Learns which models work for you | From local telemetry (runs finished, tok/s, time to first token, errors, completion acceptance), LocalPilot ranks your installed models per role. Auto routing prefers the ones that perform, and with localpilot.models.autoTune: auto (the default) it updates the default, Plan-mode, subagent, completion and image models itself, with Undo. A model you chose by hand is only replaced when it keeps failing. shows the evidence; reviews it on demand. |
Using it
- Switch Agent / Ask / Plan in the composer or with
/mode.
- Type
@ to attach context, / for commands: /help, /model, /compact, /clear, /skills, /agents, /skill <name> <task>, /skill install <owner/repo|url>, /skill browse [words], /agent <name> <task>, /context.
- Code blocks have Copy / Insert at cursor / Apply to file; approvals appear inline with Allow for session.
- Right-click in an editor → LocalPilot → Explain · Fix · Refactor · Generate Tests · Add Documentation · Fix Errors in File.
Skills, subagents and rules
.localpilot/skills/<name>/SKILL.md # name + description front-matter, Markdown instructions
.localpilot/agents/<name>.md # name, description, tools, model + system prompt body
.localpilot/commands/<name>.md # custom slash command: prompt body with $ARGUMENTS, @file, !`cmd`
.localpilot/settings.json # permission rules (allow/ask/deny) and hooks, Claude Code format
.localpilot/plugins/<name>/ # .claude-plugin/plugin.json bundling commands, agents, skills, hooks, MCP
.localpilot/rules.md · AGENTS.md · CLAUDE.md · .claude/CLAUDE.md · CLAUDE.local.md · .cursorrules · .github/copilot-instructions.md
Everything above is also read from the .claude/ folder (and ~/.claude/), so a repository configured for Claude Code works unchanged: settings.json permission rules such as Bash(npm run test:*), Edit(src/**) or Read(.env) decide what runs without a prompt, what always asks and what is refused; hooks (PreToolUse, PostToolUse, UserPromptSubmit, Stop, SessionStart, …) run your shell commands with the event as JSON on stdin and block with exit code 2; CLAUDE.md files are loaded from user, parent and project folders with @imports resolved. /permissions, /hooks, /commands and /memory show what is in force. Guide: Claude Code compatibility.
Skills you do not have yet install themselves: requests are matched against a registry of skill repositories (anthropics/skills, obra/superpowers, github/awesome-copilot by default — planning, design docs, debugging, PDF/Word/PowerPoint, architecture and more) and the best match is installed and loaded as the request starts. Settings → Skills & subagents is the management page: search the registry and install with one click, enable / disable / open / delete installed skills, see which ones get used, and set the default mode (auto / ask / off); the Skills button under each chat overrides that mode per chat and pins skills to every request. You can still install from GitHub or a URL with LocalPilot: Install Skill… (/skill install anthropics/skills/skills/pdf, localpilot skill install owner/repo). The model only sees enabled skills' names + descriptions with a rule to use those names only; a skill that clearly matches (name or trigger phrase), is pinned or forced with /skill is preloaded into the request so it is followed rather than guessed, and load_skill resolves mangled names instead of failing. Subagents run with a fresh context window and return only a final report. Full guide: Skills & subagents.
Terminal & CLI
The agent runs its commands in a visible, branded LocalPilot Agent terminal (shell integration), so you watch npm test fail and pass while the agent fixes the code. Right-click any terminal for Explain Terminal Output / Fix Failed Command, or type @terminal in the chat.
The localpilot CLI is the same agent for your shell — same tools, skills, memory, MCP servers, approvals and undo, against your local models:
localpilot "fix the failing tests" # agent mode, one task
localpilot --ask "what does main.ts do?"
localpilot # interactive session
localpilot models · localpilot doctor
Open it from VS Code (LocalPilot: Open LocalPilot Terminal — no install) or run Install 'localpilot' CLI in PATH for every terminal. Docs: Terminal & CLI.
Settings & first-run setup
LocalPilot has its own Settings screen (gear icon in the chat toolbar): a provider manager with Test connection, model and generation options, Careful / Balanced / Autonomous approval presets, inline completions, web search, memory & skills, and MCP servers — each value tagged with the scope it comes from and resettable individually. On first run a short setup wizard detects Ollama / LM Studio / llama.cpp / vLLM, recommends a model and picks an approval preset.
Configuration highlights
| Setting | Default | Purpose |
|---|
localpilot.providers | Ollama, LM Studio, llama.cpp, vLLM on localhost | Model servers (ollama or any openai-compatible) |
localpilot.defaultModel | auto | provider/model, e.g. ollama/qwen2.5-coder:14b |
localpilot.contextWindow | 32768 | Requested context (Ollama num_ctx), capped by the model; 0 = model maximum |
localpilot.approvals.* | terminal ask · write always · delete ask · network always | Autonomy level |
localpilot.parallel.maxConcurrency | 4 | Concurrent tools / subagents |
localpilot.worktrees.mode | always | Isolate each agent in its own git worktree (always / parallel / off) |
localpilot.inlineCompletion.enabled | false | Ghost-text completions |
localpilot.mcpServers | {} | MCP servers (stdio) |
All settings: Configuration.
Recommended models
| Task | Models |
|---|
| Agentic coding | gemma4:e2b/e4b/12b (recommended), nemotron-3.5-bolt:e4b, devstral, gpt-oss, mistral-small, qwen3 |
| Inline completions | starcoder2:3b, codestral, qwen2.5-coder:1.5b/7b-base |
| Screenshots / vision | gemma4, nemotron-3.5-bolt, qwen2.5vl, llava, gemma3, pixtral |
Development
npm install
npm run compile # type-check + lint + bundle
npm run smoke # offline tests (mock Ollama/OpenAI servers, agent loop)
npm test # Extension Host tests
npm run check:version # package.json, lockfile, docs config and changelog agree
npm run release -- minor # pick a specific bump; plain pushes to main auto-release a patch
Press F5 for an Extension Development Host. Architecture notes: Development.
Contributing & support
Issues and pull requests are welcome — see open issues. If LocalPilot is useful to you, a GitHub star helps others find it.
LocalPilot is built and maintained by Sathishkumar Nagarajan (GitHub: @isathish).
License
MIT © Sathishkumar Nagarajan and LocalPilot contributors.
