Open-source AI coding agent that spawns parallel agents, each bound to the LLM of your choice.
Website and documentation coming soon!
OpenDev is an open-source, terminal-native coding agent built as a compound AI system. Instead of a single monolithic LLM, it uses a structured ensemble of agents and workflows -- each independently bound to a user-configured model.
Work is organized into concurrent sessions composed of specialized sub-agents. Each agent executes typed workflows (Execution, Thinking, Compaction) that independently bind to an LLM, enabling fine-grained cost, latency, and capability trade-offs per workflow.
OpenDev automatically routes each phase of work to the right model. Five workflow slots — Normal (execution), Thinking (reasoning), Compact (context summarization), Self-Critique (output verification), and VLM (vision) — each bind independently to any LLM you configure. For example, Claude Opus handles execution, GPT-o3 handles reasoning, and a lightweight Qwen model handles compaction — all routed automatically. Together, these form a compound AI system where multiple models collaborate, each optimized for its role.
OpenDev is written in Rust — it starts in 4.3 ms, uses just 9.4 MB of memory, and ships as a single 18 MB binary. That makes it the fastest and lightest coding agent available today — up to 128x faster startup and 30x less memory than alternatives.
| Agent | Startup (mean ± σ) | Peak Memory (median) | Install Size |
|---|---|---|---|
| OpenDev 0.1.4 | 4.3 ms ± 0.4 ms | 9.4 MB | 18 MB |
| Codex 0.116.0 | 37.8 ms ± 0.8 ms (9x) | 43.7 MB (4.6x) | 116 MB |
| Claude Code 2.1.87 | 87.3 ms ± 2.0 ms (20x) | 214.6 MB (22.8x) | 188 MB |
| OpenCode 1.2.27 | 557.4 ms ± 31.8 ms (128x) | 285.9 MB (30.4x) | 90 MB |
macOS ARM64 (Apple Silicon) · Startup: hyperfine --shell=none --warmup 10 --runs 100 · Memory: /usr/bin/time -l median of 20 runs · Multipliers relative to OpenDev
- Blazing fast, ultra lightweight. 4.3 ms startup, 9.4 MB RAM, 18 MB on disk. Written in Rust with zero interpreter overhead — it launches before other agents finish loading their runtime.
- Proactive, not reactive. OpenDev can plan, execute, and iterate autonomously. Kick off a refactoring, walk away, and come back to a PR ready for review.
- Multi-provider, multi-model. Assign different models from different providers to every workflow and session, all running in parallel. Your models, your rules.
- TUI + Web UI. A full terminal UI for power users and a Web UI for visual monitoring. The Web UI supports remote sessions, so you can start a task from your phone and let OpenDev work while you sleep.
A fleet of agents, each independently exploring a different crate — all running concurrently in a single session.
Need to survey an entire codebase? Refactor across 20 crates? Run a dozen tool calls at once? Spawn a fleet.
OpenDev's agent fleet launches multiple sub-agents in parallel, each with its own LLM binding, context window, and tool access. Because the runtime is written in Rust with fully async I/O, there is zero interpreter overhead — agents fan out across your workspace and converge results back in seconds, not minutes.
You OpenDev Fleet
│ ┌─ Agent 1 → crate/agents
│ "survey all crates" ├─ Agent 2 → crate/http
│ ─────────────────────► ├─ Agent 3 → crate/tui
│ ├─ Agent 4 → crate/tools
│ ├─ ...
│ ◄── aggregated results └─ Agent N → crate/config
- Concurrent, not sequential. Every agent runs its own async task — no GIL, no queue, no waiting.
- Rust-native performance. Near-zero overhead per agent. Memory-safe parallelism via Tokio.
- Independent LLM bindings. Each agent in the fleet can target a different model or provider.
cargo install opendev-cli# Homebrew (recommended)
brew install opendev-to/tap/opendev
# Shell installer
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/opendev-to/opendev/releases/latest/download/opendev-cli-installer.sh | sh
# Or download the binary directly from GitHub Releases:
# opendev-cli-aarch64-apple-darwin.tar.xz (Apple Silicon)
# opendev-cli-x86_64-apple-darwin.tar.xz (Intel)# Shell installer (x86_64 and ARM64)
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/opendev-to/opendev/releases/latest/download/opendev-cli-installer.sh | sh
# Or download the binary directly from GitHub Releases:
# opendev-cli-x86_64-unknown-linux-gnu.tar.xz (x86_64)
# opendev-cli-aarch64-unknown-linux-gnu.tar.xz (ARM64 / Raspberry Pi)# PowerShell installer
powershell -ExecutionPolicy ByPass -c "irm https://github.com/opendev-to/opendev/releases/latest/download/opendev-cli-installer.ps1 | iex"
# Or download opendev-cli-x86_64-pc-windows-msvc.zip from GitHub ReleasesRequires Rust 1.94+.
git clone https://github.com/opendev-to/opendev.git
cd opendev
cargo build --release -p opendev-cli
# Binary at target/release/opendev (or opendev.exe on Windows)If you use the repo for development, you may also have a local symlink at ~/.local/bin/opendev pointing at target/release/opendev. That can take precedence over the Homebrew binary in /opt/homebrew/bin/opendev.
To test a Homebrew install from a clean shell state:
rm -f ~/.local/bin/opendev
hash -r
brew uninstall opendev
brew untap opendev-to/tap
brew tap opendev-to/tap
brew install opendev-to/tap/opendev
which opendev
opendev --versionSee DEVELOPMENT.md for the full local development and Homebrew testing workflow.
All release binaries, checksums, and installers are available on the GitHub Releases page.
| Platform | Architecture | Binary |
|---|---|---|
| macOS | Apple Silicon (M1+) | opendev-cli-aarch64-apple-darwin.tar.xz |
| macOS | Intel | opendev-cli-x86_64-apple-darwin.tar.xz |
| Linux | x86_64 | opendev-cli-x86_64-unknown-linux-gnu.tar.xz |
| Linux | ARM64 | opendev-cli-aarch64-unknown-linux-gnu.tar.xz |
| Windows | x86_64 | opendev-cli-x86_64-pc-windows-msvc.zip |
opendev --versionIf Homebrew reports Not a valid ref: refs/remotes/origin/main while auto-updating the tap, remove the stale local tap clone and retry:
brew untap opendev-to/tap
brew tap opendev-to/tap
brew install opendev-to/tap/opendev# Set an API key (OpenAI, Anthropic, or Fireworks -- any one will do)
export OPENAI_API_KEY="sk-..."
# export ANTHROPIC_API_KEY="sk-ant-..."
# export FIREWORKS_API_KEY="fw_..."
# Start the interactive TUI
opendev
# Or start the Web UI
opendev run ui
# Single prompt (non-interactive)
opendev -p "explain this codebase"
# Resume most recent session
opendev --continuePrefer a guided walkthrough? Run opendev config setup to interactively choose providers, models, and workflow bindings.
See the Provider Setup Guide for all 9 supported providers, authentication details, and advanced configuration.
OpenDev supports 9 LLM providers: OpenAI, Anthropic, Fireworks, Google, Groq, Mistral, DeepInfra, OpenRouter, and Azure OpenAI.
Each provider's models can be independently assigned to 5 workflow slots:
- Normal -- Primary execution model for coding tasks and tool calls
- Thinking -- Complex reasoning and planning (falls back to Normal)
- Compact -- Context summarization when history grows long (falls back to Normal)
- Critique -- Self-critique of agent reasoning (falls back to Thinking)
- VLM -- Vision/image processing (falls back to Normal if it supports vision)
Mix and match providers per slot in ~/.opendev/settings.json:
{
"model_provider": "anthropic",
"model": "claude-sonnet-4-20250514",
"model_thinking_provider": "openai",
"model_thinking": "o3"
}See the Provider Setup Guide for the full list of env vars, fallback chains, and configuration options.
OpenDev also works fully offline against local OpenAI-compatible endpoints (Ollama, LM Studio, llama-server): set OPENDEV_DISABLE_REMOTE_MODELS=1 to skip the models.dev catalog fetch and point OPENDEV_API_BASE_URL (or api_base_url in settings) at your local endpoint. See Offline and Air-Gapped Use.
Dynamic tool discovery via the Model Context Protocol for connecting to external tools and data sources.
opendev mcp list
opendev mcp add myserver uvx mcp-server-sqlite
opendev mcp enable/disable myservergit clone https://github.com/opendev-to/opendev.git
cd opendev
cargo build --workspace
cargo test --workspacecargo check --workspace # Type check
cargo clippy --workspace # Lint
cargo fmt --all # Format
cargo test -p opendev-cli # Test a specific crateDetailed local-dev, symlink, Homebrew, and release-testing notes are in DEVELOPMENT.md.
The frontend is a React/Vite app in web-ui/:
cd web-ui && npm ci && npm run buildIf you're interested in contributing to OpenDev, please open an issue or submit a pull request.
See the Roadmap for current priorities and areas where community contributions are most welcome.
- vs. Claude Code / Codex CLI / Gemini CLI: Closed-source tools that lock you into a single provider. OpenDev is fully open source and lets you mix models from any provider, independently bound per workflow (execution, thinking, critique, compaction, vision).
- vs. OpenCode: OpenCode is a great open-source coding agent with TUI, Web UI, and LSP support. However, its architecture is not modular enough to support per-workflow model binding, concurrent multi-agent sessions, or compound AI orchestration.
- vs. OpenClaw: OpenDev and OpenClaw share similar concepts around autonomous AI agents. The key difference is focus: OpenDev is purpose-built for the software development lifecycle, with context engineering, structured agent workflows, and deep code understanding.