Open source, Apache 2.0

Run a coding agent on your machine. Watch every step.

See every tool call, diff, and terminal command as it happens. Bring your own models. Keep your data local.

curl -fsSL https://openagentd.dev/install.sh | sh

shopfrontcheck_stock duplicate SKU cart lines fix

check_stock only looks at one cart line at a time, so two lines with the same SKU can both pass even when the total is over stock. Fix it, add a test that fails before the fix, and run pytest.

Grep: Searching check_stock (*.py)326ms

pattern: check_stock

Glob: Finding **/*.py286ms

pattern: **/*.py

Read: inventory.py375ms

src/shopfront/inventory.py

Read: test_inventory.py68ms

tests/test_inventory.py

Read: cart.py51ms

src/shopfront/cart.py

The cause is clear: check_stock compares each line against stock independently, so repeated SKUs never add up. I'll add the failing test first, confirm it fails, then fix.

Patch: test_inventory.py+111.8s

test_check_stock_sums_quantities_across_lines_with_the_same_sku

Shell: Confirm new test fails before fix893ms1 failed, 2 passed

DID NOT RAISE OutOfStock

The test fails as expected. Now the fix: aggregate requested quantity per SKU before comparing.

Patch: inventory.py+8-51.5s

src/shopfront/inventory.py

Shell: Run full test suite560ms10 passed

pytest -q

check_stock now sums quantities per SKU before comparing against stock, so two lines with the same SKU can no longer both pass when their combined quantity exceeds what's available.

claude-haiku-5-511s$0.0035

GitChanges (2)

No changes yet

src/shopfront/inventory.py+7-4

11 lines unchanged

14def check_stock(cart: Cart, stock: Mapping[str, int]) -> None:

15 """Raise OutOfStock for the first line item the warehouse cannot fill."""

15 """Raise OutOfStock for the first SKU whose total cart quantity exceeds stock."""

16 requested: dict[str, int] = {}

17 for item in cart.items:

17 available = stock.get(item.sku, 0)

18 if item.quantity > available:

19 raise OutOfStock(item.sku, item.quantity, available)

18 requested[item.sku] = requested.get(item.sku, 0) + item.quantity

19 for sku, quantity in requested.items():

20 available = stock.get(sku, 0)

21 if quantity > available:

22 raise OutOfStock(sku, quantity, available)

tests/test_inventory.py+11

Fig. 1A real session, replayed: the agent writes a failing test, fixes check_stock, and runs pytest.

Works with the models you already use

19+ providers, including OpenAI, AWS Bedrock, and xAI. Use your own keys.

Every step is on screen

One turn, start to finish: the agent plans, reads, edits, runs, and hands you a diff to review.

Fig. 2One turn of the agent loop. Each step has its own view in the app.

Plan
The agent writes a task checklist, so you can see what it is doing and what is done.
Read
Every tool call shows its arguments, result, and timing. Expand any of them.
Edit
File changes appear as inline diffs with real line numbers.
Run
Shell output streams into the transcript, and a real terminal sits in the same workspace.
Review
The Git panel lists every change, and language servers report diagnostics after edits.

See the work, not just the answer

Fix a bug

A whole session, start to finish: the agent reads the code, writes a failing test, patches it, and runs pytest until it passes.

See the work

Streamed replies, tool calls, and results in one view, with helper agents working beside the lead.

A lead agent working with two helpers: reports from an explorer and a researcher in the conversation, and both helpers listed under the session in the sidebar.

Review diffs

Open local folders, use git worktrees, review diffs, and get language-server diagnostics after edits.

A Git changes panel with an inline diff of src/shopfront/billing.py, removed lines in red and added lines in green, with line numbers.

Track tasks

A readable checklist shows what the agent is doing and what is done.

The Tasks panel next to a conversation, with five finished steps for adding discount codes and a 5/5 done count.

Use the terminal

A real shell in a tab next to the agent, in the same workspace.

A workspace with a terminal tab open on the right, showing git status, the last two commits, and a passing pytest run.

Everything the agent needs, in one window

Agent teams
The lead agent hands work to helper agents that run in the background, then collects their reports.
Command palette
Search app actions from the keyboard with Command K or Control K.
0 9 * * 1-5Every weekday at 09:00
1. M
2. T
3. W
4. T
5. F
6. S
7. S
Scheduler
Run a prompt on an interval, a cron schedule, or once at a set time.
Plugins
A TypeScript or JavaScript file can add a model provider or rewrite tool calls and results.
- src/
- tests/
- .env
- .env.local
- pyproject.toml
Denied paths
Block paths such as .env files from the agent's file tools with glob patterns.
mcp.json
```
{  "servers": {    "excalidraw": {      "transport": "http",      "url": "https://mcp.excalidraw.com"    }  }}
```
.openagentd/skills/
- commit/SKILL.md
- review/SKILL.md
- testing/SKILL.mdloaded
MCP servers and skills
Connect Model Context Protocol servers and add skills the agent loads when it needs them.

Fig. 3What stays on your machine, and the one thing that leaves it.

Your machine. Your keys. Your sessions.

Your prompts go to the model provider you choose. Your sessions, telemetry, and settings stay on your computer.

Local by default
Sessions, telemetry, and configuration live on your computer, and so does the API server.
Your keys, your models
Bring keys for 19+ providers, add more with plugins, or run local models with Ollama.
Open source
The app and its backend are released under Apache 2.0.

model providers: 19+
idle memory for the backend: ~20 MB
single-binary CLI: ~17 MB
native desktop app: ~26 MB

Install in one line

macOS

Terminal

curl -fsSL https://openagentd.dev/install.sh | sh

Homebrew

brew install --cask lthoangg/tap/openagentd

The desktop app runs its own backend. Add a model provider in Settings after the first launch.

Windows

PowerShell

irm https://openagentd.dev/install.ps1 | iex

The desktop app runs its own backend. Add a model provider in Settings after the first launch.

Linux

Terminal

curl -fsSL https://openagentd.dev/install.sh | sh

The desktop app runs its own backend. Add a model provider in Settings after the first launch.

iOS

SideStore source

https://github.com/lthoangg/openagentd/releases/download/latest-ios/source.json

The iOS app connects to an OpenAgentd server on your computer.

iOS install guide

CLI server

Terminal

curl -fsSL https://openagentd.dev/install.sh | sh -s -- --cli

Start the server

openagentd server start

Read the script

Terminal

curl -fsSL https://openagentd.dev/install.sh | sh

Finding the latest OpenAgentd release

Extracting the desktop app

Installing OpenAgentd.app into /Applications

Installed OpenAgentd!

Next: open OpenAgentd from Applications.

Fig. 4What the installer prints on your platform.

Questions, answered

Does my code leave my machine?

Your prompts, and the files the agent reads for them, go to the model provider you choose. Sessions, telemetry, and settings stay on your computer. With a local model through Ollama, nothing leaves it.

Which models can I use?

Any of 19+ providers, including Anthropic, OpenAI, Gemini, OpenRouter, AWS Bedrock, GitHub Copilot, and Ollama. You bring your own keys or sign in where the provider supports it, and a provider plugin can add one that is missing.

What does it cost?

OpenAgentd is free and open source under Apache 2.0. You pay your model provider for what you use, or nothing with local models.

Is there a CLI?

Yes. Install the CLI server to run openagentd server start, connect a browser or another device, or run one agent turn in the current project with openagentd run.

Can I use it from my phone?

Yes. The iOS app connects to an OpenAgentd server running on your computer, so the work still happens there.

Can I keep secrets away from the agent?

Yes. A path denylist blocks files such as .env from the agent's file tools with glob patterns. Shell commands get a best-effort scan, so keep real secrets out of reach as well.

Run a coding agent on your machine. Watch every step.

Works with the models you already use

Every step is on screen

Plan

Read

Edit

Run

Review

See the work, not just the answer

Fix a bug

See the work

Review diffs

Track tasks

Use the terminal

Everything the agent needs, in one window

Agent teams

Command palette

Scheduler

Plugins

Denied paths

MCP servers and skills

Your machine. Your keys. Your sessions.

Local by default

Your keys, your models

Open source

Install in one line

macOS

Windows

Linux

iOS

CLI server

Questions, answered