OpenAgentd
A desktop coding agent that runs on your own machine and shows every tool call, diff and terminal command.
- Source repo
- lthoangg/OpenAgentd
- Stars
- ★ 217
- Last updated
- 1d ago
- License
- Apache-2.0
- Primary language
- TypeScript
- FA score
- 53/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformChatGPT · OpenAI API · Claude API
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A solo developer who wants the agent to finish one concrete job on their laptop — such as adding rate limiting to a small Bun service and running the suite — then reviews the uncommitted diff before committing.
- Not a fit if
- Teams that need unattended CI runs on headless servers with no desktop environment
- Operators who need built-in TLS before exposing a service to the public internet
- Users who want one fixed vendor model and no bring-your-own-key setup
- Website
- openagentd.dev ↗ set on GitHub
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-20latest release v3.61.0 (2026-11-18)
- Source review
- 53/100 · Major gaps
What does this agent do, and when should you use it?
OpenAgentd is a local-first desktop app for running a coding agent on your own computer: you hand it a folder and a task, and it plans, reads, edits, runs your tests and shows every step — each tool call with its arguments and timing, every diff, and what each turn costs. It is powered by a single native Rust backend (a 46 MB binary that starts in about 50 ms), and the desktop app launches that backend itself, so sessions, settings and telemetry stay on your machine. It ships with 18 built-in model providers (Anthropic, OpenAI, Gemini, OpenRouter, Bedrock, DeepSeek and more), works with an existing ChatGPT, Copilot or Grok subscription, and runs local models through Ollama. Work is split into Plan mode, where the agent explores read-only and writes a plan you can comment on and approve, and Code mode, where you can steer, queue or stop a running turn. Beyond the desktop app, a CLI-only install provides an agentd backend and server that phone and desktop clients can connect to, and MCP servers, skills and plugins extend what the agent can reach.
After you open a folder and describe a task, the agent reads the repository and plans. In Plan mode it explores read-only and writes a plan with a "Done when" condition; you select any passage to comment on it, then Approve, or use Reply in chat to send the comments as a normal message. In Code mode it executes directly and you can steer, queue or stop the running turn. While working it calls tools to read and edit files and to run shell commands and tests; every tool call with its arguments and timing, shell output, file diffs, todos, tokens and cost streams into one transcript, which Reader mode can collapse to just the answers. A lead agent can delegate focused jobs to subagents such as explorer#1, /btw answers side questions in its own panel without entering the agent's context, and language-server diagnostics appear after edits. Repository features include Git changes, history and graph, worktrees, a real terminal and a live preview of your web app. The agentd CLI runs the backend headless — agentd server start serves the API (add --public --key --qr to pair a phone, or --cloud for remote control), and agentd run / agentd run --continue execute tasks — while the iOS and Android apps pair by QR code and show the same sessions.
- A solo developer who wants the agent to finish one concrete job on their laptop — such as adding rate limiting to a small Bun service and running the suite — then reviews the uncommitted diff before committing.
- Teams that want an approved plan before any code changes: generate the plan in Plan mode, discuss it with inline comments, then approve and let the agent work until the plan's Done when is proven.
- Users who want model costs on their own accounts: use an existing ChatGPT, Copilot or Grok subscription, or run local models through Ollama instead of committing to one vendor.
- A developer away from the desk who wants to keep following the task: pair a phone with the desktop app or an agentd server to view the same sessions, answer questions, approve plans or stop a run.
- Anyone who needs to ask a clarifying question without polluting the main context: /btw opens a separate panel whose answer never enters the agent's context.
- Teams extending the agent's reach by connecting MCP servers, adding skills and plugins, or chaining many tool calls in one scripted step.
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Install / run
curl -fsSL https://openagentd.dev/install.sh | shirm https://openagentd.dev/install.ps1 | iexbrew install --cask lthoangg/tap/openagentdcurl -fsSL https://openagentd.dev/install.sh | sh -s -- --cli
- Install notes cover
- Windows · macOS · Linux
How do you install or deploy this agent?
On macOS and Linux use the install script:
curl -fsSL https://openagentd.dev/install.sh | shOn Windows (PowerShell):
irm https://openagentd.dev/install.ps1 | iexOn macOS you can also use Homebrew:
brew install --cask lthoangg/tap/openagentdTo install only the backend, for terminal use or to serve other devices:
curl -fsSL https://openagentd.dev/install.sh | sh -s -- --cliWindows PowerShell CLI install:
& ([scriptblock]::Create((irm https://openagentd.dev/install.ps1))) -CliPlatform builds are on the releases page; unsigned builds may ask you to confirm before the first launch. Building from source is described in the Quick start section of CONTRIBUTING.md.
How do you use this agent?
Open OpenAgentd and connect a model on the first screen: paste an API key, sign in with a ChatGPT or Copilot subscription, or press Use on a key or Ollama install it finds. Then open a folder and describe what you want built; the desktop app starts its own backend, so there is nothing else to run.
Press Tab in the composer to switch between Plan and Code modes. In Plan mode, review the plan, select a passage to comment, then Approve, or use Reply in chat to send your comments as a normal message. From a terminal, use the CLI:
agentd server start
agentd run "Summarize this project"
agentd run --continue "Now add tests for it"Add --public --key --qr to server start to pair a phone, or --cloud to enable remote control; agentd --help lists every command. The server serves the API only and has no built-in TLS, so put an HTTPS reverse proxy in front of it before exposing it outside a trusted network.
What are this agent's strengths and limitations?
- Full visibility: tool calls with arguments and timing, shell output, diffs, tokens and cost land in a single transcript, and Reader mode keeps just the answers.
- Plan-then-act control: Plan mode explores read-only and produces a commentable plan; after approval the agent keeps working until the plan's Done when condition is proven.
- Provider choice: 18 built-in providers plus existing ChatGPT/Copilot/Grok subscriptions and local models via Ollama, with plugins able to add more.
- Local-first and small: one Rust backend, a 46 MB binary starting in about 50 ms, with sessions, settings and telemetry kept on the machine.
- Built for real repositories: Git changes, history and graph, worktrees, a real terminal, a live web preview and language-server diagnostics after edits.
- The same sessions extend to phones: pair by QR code and reply, approve or stop the agent from iOS or Android, with optional remote control.
- Model usage is not included: you need your own provider key, subscription or local Ollama install, and costs depend on the model you choose.
- The CLI server has no built-in TLS; you must deploy an HTTPS reverse proxy before exposing it beyond a trusted network.
- Unsigned builds may trigger OS confirmation on first launch, which adds friction for managed or enterprise distribution.
- Mobile use depends on pairing with the desktop app or configuring remote control; it is not a standalone cloud service.
- Automatic work after plan approval depends on the plan's stated Done when condition, so plan quality directly governs how safe the unattended run is.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| OpenAgentd This agent | 53 · Major gaps | Desktop appFree + model costs | ★ 217 | 1d ago | TypeScript | ChatGPT · OpenAI API · Claude API |
| Ogcode | 50 · Major gaps | CLIFree + model costs | ★ 148 | 3d ago | Go | OpenAI API · Claude API |
| Pi Agent Desktop | 78 · Good | — | ★ 353 | 9d ago | TypeScript | OpenAI API |
| Synara | 57 · Major gaps | Desktop appFree + model costs | ★ 2.3k | 1d ago | TypeScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
SECURITY.md and AGENTS.md document constant-time key comparison, network guard, workspace validation, path-traversal protection, argument-list subprocesses, 0600 atomic credential writes, and request size caps, giving solid static evidence for least_privilege/user_confirmation/data_flow_transparency/sensitive_data_handling at 2; deductions for plaintext credentials without OS keychain and a lexical-only (non-sandbox) shell check. dependency_security gets 2 for cargo audit and locked builds in cloud.yml, but only three Rust lockfiles are audited and web/Bun dependencies are not covered. external_effects gets 2 for Plan mode and pre-commit review, though AGENTS.md lets the agent commit locally (reversible). rollback gets 1: only git commits and update verification are shown, no session/file-level rollback. source_attribution gets 1: README and LICENSE only, no NOTICE or third-party attribution.
self_consistency gets 2: README, AGENTS.md, and SECURITY.md agree on architecture and security boundaries. dependency_availability gets 1: Bun 1.4.0, Rust, cargo-nextest, python3, lsof/nc are listed but without lockfile detail or fallback paths. failure_messages gets 1: only drive.sh doctor and scattered code comments show error handling conventions; no systematic failure-message standard.
audience_and_scenarios gets 2: desktop, CLI, mobile, and multiple providers are clearly covered. capability_boundaries gets 2: Plan/Code modes, read-only exploration, and approval flow define boundaries. trigger_precision gets 1: SKILL.md descriptions are broad (e.g. guidelines triggers on writing/reviewing/refactoring code) without precise triggers. environment_fit gets 2: macOS/Linux/Windows install paths and Tauri system dependencies are stated.
information_architecture gets 2: README, layered AGENTS.md instructions, and documents/docs/features.md index. install_notes gets 2: curl, brew, PowerShell, and CLI install paths. naming_stability gets 1: no versioned naming or deprecation policy shown. examples_and_faq gets 1: only a quick tour with screenshots, no FAQ. known_limitations gets 2: SECURITY.md and README disclose no built-in TLS, non-sandbox shell check, plaintext credentials. license gets 2: full Apache-2.0 text. versioning_changelog gets 1: releases and version-cited features exist but no in-repo CHANGELOG. maintenance_responsibility gets 1: publisher unverified and SECURITY.md only promises best-effort response.
output_usability gets 2: transcript, diffs, cost display, and Reader mode improve usability. marginal_value gets 2: local-first, multi-provider, mobile, and observability differentiate it from generic coding agents. cost_benefit gets 1: 46MB/50ms claims lack benchmark evidence and users must supply model keys, so cost-benefit is asserted rather than shown.
claim_traceability gets 1: many README claims point to SECURITY.md and contract docs, but performance claims (46MB, 50ms) have no supporting evidence. cross_source_corroboration gets 2: README, SECURITY.md, AGENTS.md, and CI workflows corroborate security and testing claims. fact_inference_separation gets 1: docs blend philosophy with facts and do not clearly separate inference from verified fact.
- Credentials are stored plaintext with 0600 in credentials.yaml, no OS keychain; the shell file guard is lexical, not an OS sandbox, so agent commands can still reach sensitive files indirectly.
- The CLI server has no built-in TLS; an HTTPS reverse proxy is required before exposing it beyond a trusted network.
- Publisher identity is unverified and maintenance is best-effort only; enterprises should verify before adoption.
- Dependency auditing covers only three Rust lockfiles; web/Bun dependencies are unaudited, and performance claims (46MB/50ms) lack benchmark evidence.
- AGENTS.md lets the agent commit locally; reversible but teams should confirm they accept automated commits.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
Release notes excerpt
What's changed - Remote control on Android. The Android app (Android 11 or later, with a screen lock) can sign in to your account or join it by scanning a code, add and approve devices, and open your computers through your account or directly on your home network, as the iPhone app does. Its keys stay in the Android Keystore, and approving, adding or removing a device asks for your fingerprint or screen lock. Done, Needs you and Failed notifications from your other computers arrive as pushes. …View on GitHub ↗ -
Release notes excerpt
What's changed - A cleaner Connectors page, with logos. Settings → Connectors now lists your own connectors first, each with the service's logo, its status, a switch and Connect, followed by a compact gallery of the ones you have not added, under one search box. Added connectors no longer show twice. The composer's "+" menu shows full names, a status line and one-line errors (the full text on hover), and says "Connect" instead of "Sign in". …View on GitHub ↗ -
Release notes excerpt
What's changed - Keys without exposing them. When a task needs an API key or token, the agent asks for it on a card. You paste the value once, on your computer or your phone, and choose which projects may use it. The agent, the transcript and the logs only ever see [secret NAME], and two projects can keep different values under the same name. Settings → Secrets lists what is stored. - Plan comments go to the chat. In a plan review, Reply in chat moves your comments into the message box, so your feedback shows in the conversation. The Plan panel's separate feedback box is gone. …View on GitHub ↗
FAQ
Do I have to pay for OpenAgentd itself?
Does my code or session data leave my machine?
Can I stop the agent while it is running commands?
How do I run it headless on a server?
Can I wire in my own tools?
Does OpenAgentd work on Windows?
irm https://openagentd.dev/install.ps1 | iex.Is OpenAgentd open source?
Is OpenAgentd safe to use?
Is OpenAgentd still maintained?
Who makes OpenAgentd?
What is OpenAgentd's official website?
Thanks — we use this to improve the page.