Rabbitrabbit/ide
The AI coding IDE · v1.35.40 · MIT

Built for developers who ship with local models, hybrid cloud routing, and an MCP token economy that slashes context spend. Local-first by default — telemetry off by default, no cloud code index, account optional. No unsolicited lectures. Just code.

Download for macOSSee it in action
macOSLinuxWindowsarm64 + x64

Trusted by developers who build with every major model and open standards

VS Code OSS
MCP
Anthropic
OpenAI
Ollama
Llama
Qwen
DeepSeek
Mistral
Tree-sitter
LSP
Tauri-ready
VS Code OSS
MCP
Anthropic
OpenAI
Ollama
Llama
Qwen
DeepSeek
Mistral
Tree-sitter
LSP
Tauri-ready
Token economy

Stop renting tokens you already paid for.

Rabbit's MCP layer (stoken) caches reads, compresses output, and prunes context with a project-wide dependency graph. The result: the same task, a fraction of the tokens.

ctx_read cache

cache hits

Re-reading the same file costs ~13 tokens instead of 500+. Cache survives sessions.

signatures / map modes

smaller reads

Read only the AST skeleton when you don't need the full file. Imports, exports, types.

ctx_shell compression

compressed

Strips ANSI, deduplicates noise, keeps essentials. 90+ commands recognized.

graph-pruned context

pruned

Dependency graph trims irrelevant files before they touch the prompt.

Illustrative example

“Refactor auth module + write tests”

Context sent

less

Typical AI IDE140k tokens
Rabbit (stoken)35k tokens
$ ctx_read src/auth.ts → full | 1.2k tokens
$ ctx_read src/auth.ts → cached | 13 tokens
$ ctx_read src/auth.ts mode=signatures → 87 tokens
Local models

Run frontier open weights. On your machine.

Local inference is a first-class citizen in Rabbit, not a fallback. Point it at any OpenAI-compatible endpoint — Ollama, llama.cpp, vLLM, LM Studio, or MLX — and the full agent just works: tool calling, streaming, and whole-project context, with zero telemetry. Your GPU, your weights, your rules.

GLM-4.7-Flash

Zhipu

30B · lightweight coding

ollama pull glm-4.7-flash

Qwen3-Coder

Alibaba

30B-A3B / 480B-A35B

ollama pull qwen3-coder:30b

DeepSeek-V3.1

DeepSeek

671B · MoE, 37B active

ollama pull deepseek-v3.1

Devstral

Mistral

24B · agentic coding

ollama pull devstral

Codestral

Mistral

22B · code generation

ollama pull codestral

Llama 4

Meta

Scout 109B / Maverick 400B

ollama pull llama4:scout

Gemma 3

Google

1B / 4B / 12B / 27B

ollama pull gemma3:27b

Phi-4

Microsoft

14B · dense reasoning

ollama pull phi4

Air-gapped by default

Pick a local model and your code never leaves the machine. Telemetry is off, there's no cloud code index — nothing to leak.

Zero rate limits

No quotas, no throttling, no per-token meter. Push as many tokens an hour as your GPU can serve.

Bring your own runtime

Ollama, llama.cpp, vLLM, LM Studio, MLX — if it speaks the OpenAI-compatible API, Rabbit's full agent drives it.

Apple Silicon friendly

Point Rabbit at any MLX or llama.cpp Metal server with an OpenAI-compatible endpoint — your M-series GPU runs inference, Rabbit runs the agent loop.

Cloud routing

Route to frontier models. In the cloud.

When you want maximum capability, Rabbit routes to hosted frontier models over your own keys or Rabbit Cloud — same agent, same tools, same whole-project context. Local stays local; the cloud is one capability-aware hop away.

Claude Opus 5.5

Anthropic

Deep reasoning · hardest agentic work

Routed · cloud

Claude Sonnet 5

Anthropic

Opus-class coding at Sonnet speed

Routed · cloud

GPT-5.6

OpenAI

General frontier · tops Terminal-Bench

Routed · cloud

Gemini 3.5 Pro

Google

2M context · multimodal Deep Think

Routed · cloud

DeepSeek V4 Pro

DeepSeek

Cost-efficient MoE frontier · 1M context

Routed · cloud

Grok 4.7

xAI

Real-time knowledge · agentic tool use

Routed · cloud
Hybrid routing

Cheap locally. Smart in the cloud.

Rabbit routes each agent step to the right model. Plan and search locally for free. Send only the hard codegen step to a frontier model — Claude Opus 5, GPT-5.6, or Gemini 3.5 Pro. Your wallet feels the difference immediately.

  • Capability-aware policy — tool-calling work never lands on a text-only model.
  • Fallback chains — local → mid-tier cloud → frontier.
  • Execution preference (auto / cloud / local) and prefer-provider hints.
  • A daily budget warning keeps spend visible; hard caps apply to agent spawns.
  • Outsource entire tasks via the `rabbit-cli` headless runner.

routing policy

rabbit-ai.routing.execution       auto      # auto · cloud · local
rabbit-ai.routing.preferProvider  auto      # or pin: anthropic · openai …

# capability-aware fallback chain
  → local      llama.cpp / Ollama                                     · free
  → mid-tier   Claude Haiku 4.5 · GPT-5.4 mini · Gemini 3.1 Flash Lite · cheap
  → frontier   Claude Opus 5 · Sonnet 5 · GPT-5.6 · Gemini 3.5 Pro     · hard steps

rabbit-ai.budget.dailyUsd         5.00      # warns — never a hard cutoff
01
Local · free
02
Mid-tier cloud
03
Frontier · hard steps
No moral filters

The model writes code. Not lectures.

Rabbit is built for adults shipping production software. The default agent prompt skips the apology preamble and unsolicited lectures — and you own it outright. Swap in your own instructions, pair Rabbit with a local model, and your loops run fully uncensored end to end.

  • No “I'm sorry, but I can't” boilerplate.
  • No unsolicited security/ethics monologues.
  • No silent edits “for your own good”.
  • Pair with any local model for fully uncensored loops.
I cannot help with that request.stripped
As an AI assistant, I must decline...stripped
I'm not able to write code that...stripped
This could be used for harmful purposes...stripped
Have you considered the ethical implications?stripped
Here is the diff. Tests are passing.shipped
Privacy · local-first

The IDE that doesn’t watch you.

Rabbit is local-first by design. Telemetry off by default, no cloud code index, no prompt logging. Cloud features are opt-in and only touch what they need.

Telemetry off by default

Analytics and crash reporting ship off and stay opt-in — the editor sends nothing unless you turn them on. Usage is recorded only when you opt into Rabbit Cloud routing, and only to bill it.

No account required

Download. Run. Code. An account is only needed for the optional cloud proxy, subscriptions, and auto-updates.

Fully offline-capable

Pair with a local model and pull the network cable. Agent, autocomplete, search, MCP, refactors keep working.

Your keys, your machine

Provider API keys are stored locally in your settings and sent only to the provider you chose. Never synced to our servers, never logged.

Your code stays local by default

Indexing, memories, sessions live in ~/.rabbit-ide/. Code is sent only to the model endpoint you configured — local or cloud.

No vendor lock-in

Open formats for memories, rules, sessions. Walk away whenever you want — everything is plain JSON on your disk.

Our pledge

Collect the minimum. Keep it honest.

We can’t sell what we don’t collect. If you never sign in, we hold nothing about you. If you do, we store only your account, balance and usage — nothing from your codebase.

# default privacy posture — every install

[telemetry]
enabled       = false   # off by default — opt-in
crash_reports = false   # off by default — opt-in

[cloud]
account_required = false  # optional, for proxy/updates
prompt_logging   = false  # prompts are not stored
code_upload      = false  # no cloud code index

[storage]
location = "~/.rabbit-ide/"
sync     = "off"          # opt-in only
Agent · MCP native

An agent with 58 native tools. Plus every MCP server you bring.

Rabbit ships with 58 built-in tools across 8 groups — read the repo, edit with review, run a shell, trace a dependency graph, keep context cheap, remember across sessions, browse the web, and measure every token. Drop any MCP server into ~/.rabbit-ide/mcp-servers.json and its tools appear in the next turn — no restart, no plugin install.

8 families · 58 native tools

Read & navigate

Explore the codebase without burning tokens on full-file re-reads.

7 tools

Read & navigate

read_file

Cached read · full / signatures / map / lines

Signature and map modes return AST skeletons instead of whole files. Cache hits cost ~13 tokens vs hundreds on repeat reads.

How it works

Reads any workspace file through the stoken cache layer. Modes: full (raw text), aggressive (code with comments and blank lines stripped — same semantics, fewer tokens), signatures/map (top-level symbols only), summary, lines:N-M (range), toon (lossless TOON encoding for JSON, ~30–60% savings), and auto (picks the best mode per file type). Prefer aggressive or auto for large code files; use toon/auto for JSON configs. Re-reads after edits are cheap cache hits (~13 tokens). Supports fresh=true to bypass stale cache.

~/.rabbit-ide/mcp-servers.json

live
{
  "stoken":    { "command": "stoken",    "args": ["serve"] },
  "postgres":  { "command": "mcp-pg",    "args": ["--url", "$DATABASE_URL"] },
  "puppeteer": { "command": "mcp-pptr",  "args": [] },
  "figma":     { "command": "mcp-figma", "env": { "FIGMA_TOKEN": "$FIGMA_TOKEN" } }
}

One JSON file. Any MCP server listed here is wired into the next agent turn automatically — stdio, HTTP, or your own binary.

drop-in

4 shown · unlimited

  • stoken

    stoken serve

    preinstalled

  • postgres

    mcp-pg --url $DATABASE_URL

    query your DB

  • puppeteer

    mcp-pptr

    drive a browser

  • figma

    mcp-figma

    read design files
  • + your own MCP server

Tools from every listed server appear in the agent's prompt automatically — no plugin install, no restart.

Compatibility

Drop in. Nothing to relearn.

A VS Code OSS fork — your extensions, keybindings, settings, snippets, and workspace files load unchanged. On first open, Rabbit auto-imports AGENTS.md, .cursorrules, .devin/, and MCP configs from other AI editors.

VS Code extensionsInstalled from OpenVSX by default, with Marketplace fallback where licensing permits.
.cursorrulesLoaded from repo root on first open as persistent chat and agent rules.
.devin/ configWorkflows, hooks, and permissions parsed from the directory and mapped to Rabbit equivalents.
AGENTS.mdRepo-level agent instructions loaded on open, following the shared cross-editor spec.
MCP serversPer-workspace and global registry; servers inherited from other AI editors' configs.
LSP / DAPLanguage servers and debug adapters inherited from the OSS base — every extension works.
Settings syncSelf-hosted sync server, opt-in — your settings and keybindings never leave your infrastructure.
TelemetryFully off at install; opt-in only, with no silent data collection.
—
rabbit-cli

Same agent. In your terminal.

rabbit-cli brings the agent to your terminal — headless, scriptable, CI-friendly. A focused subset of the agent's tools (read, search, edit, shell, web), driving any provider you point it at.

use

CI guards · code review bots

use

Migrations · refactors at scale

use

Outsourced tasks · overnight runs

use

SSH into a remote box, run it there

~ / acme-monorepo
zsh
$ npm i -g rabbit-cli
$ rabbit config              # pick provider + model
$ rabbit "bump every workspace to node 22 and fix
          every type error, then open a PR"

  ▸ scanning 142 packages
  ▸ [search] [read_file] [edit_file] [shell]
  ▸ 38 files changed · 0 type errors · 312 tests green
  ✓ opened PR #4821
Pricing

Free for life if you want. Cheap if you don’t.

Rabbit is free to download and free forever for local models. Cloud routing is optional, transparent, and capped.

Free

$0forever

Everything local. Bring your own GPU or keys. Unlimited local tokens.

  • Full IDE + agent
  • Unlimited local models (Ollama, llama.cpp, MLX)
  • BYO Anthropic / OpenAI / DeepSeek keys
  • 50 requests / day
  • 200K tokens / month
  • 4 cloud models (Claude Haiku, DeepSeek, GPT-5.4 mini, Gemini Flash Lite)
  • Community support
Download
Most popular

Pro

$20/ month

Cloud routing without your own keys. Higher limits, all models.

  • 500 requests / day, then usage-based
  • 5M tokens / month
  • All cloud models (Claude Opus, Sonnet, GPT-5, Gemini 3.5 Pro…)
  • Cloud agents
  • Advanced analytics & cost tracking
  • Priority support
Start Pro

Enterprise

Customcontact us

For teams that run the agent all day, every day.

  • Everything in Pro
  • Custom model hosting
  • SSO / SAML
  • Dedicated support & SLA guarantee
  • Custom integrations
  • Self-host & air-gapped deployment
Contact Sales
FAQ

Questions, answered.

Three things. First, the MCP-powered token economy: cached reads, signature/map modes, compressed shells. Second, first-class local models — not an afterthought, but the default. Third, you own the system prompt: the agent skips unsolicited preamble and lectures, and paired with a local model you get fully uncensored loops.

The new way to build software — local models for free, cloud when you need power, and a token economy that cuts context spend.

Download RabbitSee pricing