Acorn

An autonomous coding agent that lives in your terminal.

Acorn reads your code, writes files, runs commands, and refactors across your entire codebase — powered by Google's Gemini.

Website: acorncli.dev

     █████╗  ██████╗ ██████╗ ██████╗ ███╗   ██╗
    ██╔══██╗██╔════╝██╔═══██╗██╔══██╗████╗  ██║
    ███████║██║     ██║   ██║██████╔╝██╔██╗ ██║
    ██╔══██║██║     ██║   ██║██╔══██╗██║╚██╗██║
    ██║  ██║╚██████╗╚██████╔╝██║  ██║██║ ╚████║
    ╚═╝  ╚═╝ ╚═════╝ ╚═════╝ ╚═╝  ╚═╝╚═╝  ╚═══╝

Features

• Real-time streaming — tokens appear as they're generated, not after
• Smart model routing — scores each request on size, scope, and intent to pick Flash or Pro
• Persistent shell — cd, export, and venv activation carry across commands
• Web search built in — searches and reads pages with no API key required
• MCP support — plug in any Model Context Protocol server for extra tools
• Colored diffs — every file change shown as a real diff with line numbers
• Surgical file editing — modifies specific lines, not entire files
• Multi-file refactoring — reads and edits across your whole codebase
• Image/screenshot analysis — attach images for the model to analyze
• Auto-retry with error recovery — adapts when things fail
• Session persistence — resume conversations where you left off
• Undo support — revert the last file change instantly
• Cost tracking — see what you're spending per session
• Project config — per-repo .acorn.toml for custom settings
• Project instructions — per-repo .acorn.md for custom context
• Permission system — 3-tier safety (safe/ask/deny)
• Git-aware — understands your branch, status, and project structure
• Dual auth — works with simple API key or Vertex AI (enterprise)
• Update notifications — know when a new version is available

---

Quick Start

# Install from PyPI
pip install acorn-agent

# Run it
acorn

On first run it will ask you to choose authentication:

1. Gemini API key — get one free at aistudio.google.com (easiest)
2. Vertex AI — use your GCP project (enterprise)

Or set it via environment variable:

export GEMINI_API_KEY="your-key-here"
acorn

---

Installation (Full Guide)

Prerequisites

Requirement	Why
Python 3.11+	Runtime
Gemini API key or GCP project	Model access

Step 1: Install

pip install acorn-agent

Or from source:

git clone https://github.com/andamagodwin/acorn.git
cd acorn
pip install -e .

Step 2: Authentication

Option A: API Key (easiest)

Get a free API key at aistudio.google.com/apikey, then:

export GEMINI_API_KEY="your-key-here"
acorn

Or just run acorn and it will prompt you on first launch.

Option B: Vertex AI (enterprise)

For organizations using GCP:

gcloud auth application-default login
gcloud services enable aiplatform.googleapis.com
export ACORN_PROJECT="your-gcp-project-id"
acorn

Step 3: Run

cd ~/your-project
acorn

---

Usage

# Start in current directory
acorn

# Use a specific model
acorn --model gemini-2.5-flash

# Use an API key directly
acorn --key YOUR_API_KEY

# List available models
acorn --models

# Disable streaming
acorn --no-stream

# Disable smart routing (always use Pro)
acorn --no-routing

# Auto-approve all file writes (careful!)
acorn --unsafe

# Override GCP project (Vertex AI mode)
acorn --project my-project-id

# Help
acorn --help

Attach Images

Just include an image path in your message:

> what's wrong in screenshot.png
> implement the design in mockup.jpg

Gemini will analyze the image and respond accordingly.

---

Commands

Command	What it does
/help	Show all commands
/model <name>	Switch model (or list available)
/cost	Show session cost breakdown
/status	Token usage, routing stats, cost
/plan	Show current task execution plan
/undo	Revert the last file change
/clear	Reset context and session
/sessions	List saved sessions
/config	Show current configuration
/exit	Quit

Tool commands

Command	What it does
/routing	Explain why the last request went to Flash or Pro
/routing on|off	Toggle smart routing
/shell	Show shell mode and current working directory
/shell reset	Clear shell state (cd, exports, venv)
/mcp	List MCP servers and the tools they expose
/web on|off	Toggle web search and page fetching

---

Persistent Shell

Commands run in one long-lived shell session, so state carries between calls:

> set up the project

~ Running python -m venv .venv
~ Running source .venv/bin/activate     <- still active below
~ Running pip install -r requirements.txt
~ Running pytest                        <- runs inside the venv

Without this, every command starts from scratch and the agent has to chain everything with &&. Run with --no-shell to go back to one process per command.

A timed-out command is interrupted on its own — the session and its state survive. Interactive programs (vim, less, top, a bare python) are refused up front, since they would hang the session waiting for input.

---

Web Search

Acorn can search the web and read pages without any API key or signup:

> what changed in the latest fastapi release?

~ Searching for 'fastapi latest release changelog'
~ Fetching fastapi.tiangolo.com/release-notes/

Two tools are available to the model: web_search for finding pages and fetch_url for reading one. Disable both with --no-web or /web off.

---

MCP Servers

Acorn speaks the Model Context Protocol, so you can plug in any MCP server and its tools become available to the agent.

Configure servers in ~/.acorn/mcp.json (global) or .mcp.json (per project):

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"]
    },
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_TOKEN": "ghp_..." }
    }
  }
}

This is the same config format other MCP clients use, so existing files work as-is. Tools appear to the model as mcp__<server>__<tool>, and since they run third-party code they prompt for permission before each call.

Run /mcp to see what's connected. A server that fails to start is reported there and never blocks startup. Use --no-mcp to skip them entirely.

---

Project Configuration

Drop a .acorn.toml in any project root to customize Acorn's behavior for that repo:

[model]
pro = "gemini-2.5-pro"
flash = "gemini-2.5-flash"
temperature = 0.2
max_output_tokens = 65536

[routing]
enabled = true
threshold = 200
# Let Flash break ties when the heuristics are unsure (one cheap extra call)
classifier = true

[shell]
# Share cwd, exports, and venv activation across commands
persistent = true

[web]
enabled = true

[mcp]
enabled = true

[project]
gcp_project = "your-project-id"
location = "us-central1"

[permissions]
safe_commands = [
    "npm run", "npm test", "cargo build",
    "python -m pytest", "make",
]

Acorn auto-detects this file by walking up from your working directory.

Project Instructions (.acorn.md)

Drop a .acorn.md in your project root to give Acorn persistent context about your project:

# Project: MyApp

## Tech Stack
- Python 3.12, FastAPI, SQLAlchemy
- Frontend: React + TypeScript

## Conventions
- Use snake_case for Python, camelCase for TypeScript
- Always add type hints
- Tests go in tests/ with pytest

## Important
- Never modify migrations directly — use alembic
- The auth module is being rewritten, don't touch auth/legacy/

This gets injected into the system prompt, so Acorn always knows your project's conventions.

---

For Collaborators

Joining the Project

1. Get added as a collaborator on the GitHub repo
2. Clone and install:

   git clone https://github.com/andamagodwin/acorn.git
   cd acorn
   pip install -e .
   

3. Set up GCP authentication:

   gcloud auth login
   gcloud auth application-default login
   

4. Ask the project owner to add you to the GCP project with the "Vertex AI User" role in IAM
5. Run acorn from any directory

IAM Setup (For Project Owner)

To add a collaborator to your GCP project:

gcloud projects add-iam-policy-binding YOUR_PROJECT_ID \
  --member="user:their-email@gmail.com" \
  --role="roles/aiplatform.user"

Or in the console: IAM & Admin > IAM > Grant Access > Role: "Vertex AI User"

---

Architecture

acorn/
├── config/
│   ├── settings.py        — Models, permissions, safety rules
│   └── project_config.py  — .acorn.toml loader
├── core/
│   ├── agent.py           — Main brain: streaming agentic loop
│   ├── context.py         — Context window with auto-compaction
│   ├── costs.py           — Token usage and cost tracking
│   ├── planner.py         — Multi-step task planning
│   ├── router.py          — Signal-scored model routing (Flash vs Pro)
│   └── session.py         — Session persistence to ~/.acorn/sessions/
├── tools/
│   ├── filesystem.py      — Read, write, edit, search, list, diffs
│   ├── terminal.py        — Command execution with timeout
│   ├── shell.py           — Persistent shell session
│   ├── web.py             — Web search and page fetching
│   ├── mcp.py             — Model Context Protocol client
│   └── git_tools.py       — Git-aware project understanding
├── ui/
│   └── terminal_ui.py     — Terminal UI with markdown rendering
└── main.py                — CLI entry point

---

How It Works

1. You type a message
2. Acorn scores it on size, scope, and intent, then routes to Flash or Pro
3. The model calls tools — reads files, edits code, runs commands, searches the web
4. Tool results feed back for up to 25 iterations per turn
5. If something fails, it sees the error and adapts
6. Response streams to your terminal in real-time

Routing

Rather than matching keywords, each request is scored on independent signals: length, pasted code or stack traces, how many files it names, whether it asks for a change or an explanation, and whether something is described as broken. /routing shows the score and every signal behind the last decision.

Keyword matching gets this wrong in both directions — "what is wrong with my entire auth system?" opens like a definition question but is real debugging work, while "refactor auth" is only 14 characters. Scoring separates them.

Requests that land in the ambiguous middle are settled by one cheap Flash call. That's a round trip, so it only happens when the signals genuinely conflict.

---

Safety

Layer	What it does
Blocked commands	rm -rf /, mkfs, fork bombs — permanently blocked
Safe commands	ls, git status, grep — auto-execute
Permission prompts	File writes, unknown commands — asks you first
File backups	Every edit is backed up for /undo
Context isolation	Won't touch .env files or expose secrets

---

Cost

Model	Input	Output	Typical session
Gemini 3.1 Pro	$1.25/M tokens	$10.00/M tokens	$0.50–$2.00
Gemini 2.5 Flash	$0.15/M tokens	$0.60/M tokens	$0.02–$0.10

Smart routing saves ~70% by using Flash for greetings, questions, and simple lookups.

Use /cost to see your spending mid-session.

---

Troubleshooting

Problem	Fix
ModuleNotFoundError: No module named 'acorn'	Run pip install -e . from the acorn directory
404 NOT_FOUND model error	Gemini 3.1 requires location="global" (already set). If using 2.5 models only, you can use us-central1
Permission denied on gcloud	Run gcloud auth application-default login
Vertex AI API not enabled	Run gcloud services enable aiplatform.googleapis.com
Session won't resume	Delete ~/.acorn/sessions/ and restart

---

License

MIT