Plan, edit, verify
Reads your codebase, plans steps, picks the best model, edits files, runs tests, and self-reviews before declaring anything done.
v1.0.17 · MIT · Built in Rust
The agentic CLI
for your terminal.
Alphacode reads your repo, picks the right model, edits files, runs tests, and ships — with your permission at every risky step.
Works with
Claude
GPT-4o/5
Gemini
DeepSeek
Mistral
xAI
Groq
Moonshot
Ollama
+45 more
~/projects/horizon
sonnet-4.5
running
0.0s
0/42 tools
21
GitHub stars
40+
Built-in tools
50+
Model providers
0
Telemetry by default
MIT
License
Everything an agent should be.
Any model, any time
Claude, GPT, Gemini, Copilot, Cursor, Bedrock, OpenRouter, or any OpenAI-compatible endpoint. Switch mid-conversation with Ctrl+T.
Swarm mode
Big jobs split into independent pieces, run by multiple agents in parallel, then merged and reviewed automatically.
Safety built in
Blocks catastrophic commands outright. Confirms before anything risky. Never sends your code anywhere by default.
40+ tools
File editing, regex/AST search, shell, browser control, web fetch, memory, scheduling, diagrams, PDFs. No plugin surgery.
Never lose work
Sessions saved to disk. alphacode --resume reopens exactly where you left off — even after a crash.
Big jobs, run in parallel.
A planner agent decomposes your goal into independent DAG nodes. Multiple agents execute them concurrently — each with the right model — then results merge and self-review.
Bring your own key. Use any provider.
First-class for the big ones, OpenAI-compatible for the long tail. Anything that speaks HTTP + JSON works.
Anthropic / Claude
Anthropic API
OpenAI
OpenAI API
Google Gemini
Gemini API
GitHub Copilot
Cursor
AWS Bedrock
Azure OpenAI
OpenRouter
DeepSeek
xAI (Grok)
Mistral
Groq
Perplexity
Together AI
Fireworks
Deep Infra
Hugging Face
Cohere
Moonshot AI · Kimi
Kimi Code
Z.AI
Alibaba Cloud Coding Plan
Xiaomi MiMo
MiniMax
Antigravity
Ollama
LM Studio
OpenAI-compatible (any)
OpenCode Zen
OpenCode Go
Baseten
Cerebras
NVIDIA NIM
Chutes
Nebius Token Factory
Scaleway
STACKIT
Cortecs
Comtegra GPU Cloud
FPT AI Marketplace
Firmware
302.AI
GMI Cloud
ZeroG
Celeris
AgentRouter
TokenRouter
Dragonmeta
Alphacode Subscription
Google / Gmail
Auto Import
Don't see yours? Point api_base at any OpenAI-compatible endpoint and you're done.
One line. Verified, then on PATH.
$ curl -fsSL https://raw.githubusercontent.com/dragonked2/alphacode/main/scripts/install.sh | bashPS> iwr -useb https://raw.githubusercontent.com/dragonked2/alphacode/main/scripts/install.ps1 | iexgit clone https://github.com/dragonked2/alphacode.git
cd alphacode
cargo build --release
Then verify:
1which alphacode
2alphacode --version
3alphacode doctor
4alphacode
Lean by design. Measured, not just claimed.
A coding agent can look fine with one session and balloon when you actually use it. RAM snapshots for one and ten concurrent sessions.
RAM · 1 active sessionMB
RAM · 10 active sessionsMB
Questions, answered.
Is Alphacode really free?
Yes. MIT-licensed. You pay for the model API you point it at (or run a local one). No telemetry, no accounts, no upsells.
Does it send my code anywhere?
Only to the model provider you configure. The CLI is local. Nothing leaves your machine unless a model call needs context.
Which models work?
Any OpenAI-compatible HTTP endpoint. Tested with Claude (Anthropic), GPT-4o/5 (OpenAI), Gemini, Copilot, Cursor, Bedrock, and OpenRouter.
How does swarm mode work?
A planner turns your goal into a task DAG. Workers — each with their own model — execute independent nodes in parallel. A reviewer node merges and verifies the diff.
Can I use it on a remote box with no GUI?
Yes. Pure terminal. Pipe-friendly. Plays well with tmux, screen, and SSH.
How is this different from Claude Code or Codex CLI?
It's MIT, model-agnostic, swarm-capable, and Rust-fast. Pick the comparison that matters to you.
Try it on a real repo. It only takes a minute.
Open any project, run alphacode, and describe what you want changed.