by Ibrahim (ibmshaikh)
Open-source desktop agent that runs your own LLM on files, shell and web.
Worth a trial for developers who hold an API key or run local models and want a free chat-first desktop agent with no account. It replaces a paid agent subscription only if you accept a months-old project with one maintainer and a last release dated 12 August 2026.
A no-account desktop agent harness that pairs any OpenAI-compatible model with file, shell, MCP and skill tools.
There is little lock-in because the app is MIT licensed and holds no account, so its value depends on the polish of its desktop harness.
Main trade-off
A young single-maintainer project with few users, no recent release and conflicting statements about telemetry.
The app is free and MIT licensed, with no account or subscription. You pay your own model provider for tokens, and optional web search uses a TinyFish key.
| System | Type |
|---|---|
| Ollama | API |
| LM Studio | API |
| OpenRouter | API |
| TinyFish | API |
OpenPilot is an open-source desktop AI agent that connects to any OpenAI-compatible model and uses file, terminal, web, MCP and skill tools to finish coding tasks. It is semi-autonomous, asking approval for sensitive actions, runs locally on Windows, and its repository has 21 GitHub stars.
Bring your own LLM
Works with any OpenAI-compatible endpoint, from a cloud API to OpenRouter, Ollama or LM Studio, with several models switchable per chat.
File tools in a workspace
The agent reads, writes, edits, greps and globs inside the folder you point it at, and shows each change as a tool card.
Live terminal
Runs real shell commands such as installs and test suites and streams stdout back, with approve or deny for sensitive actions.
Optional web research
With a TinyFish key the agent can search and fetch pages, and without one it still works on local files and the terminal.
MCP, skills and memory
MCP servers toggle per chat, skills load from a slash command in the composer, and AGENTS.md carries project preferences.
In-app token tracking
Shows session and lifetime token usage by model so you can see what your own API key is spending.
| Agent | Where it runs | Price | Maturity | Rating |
|---|---|---|---|---|
| Your machine | Free | Beta | 3.3 | |
| Your machine · API | Free plan | Available | -- | |
| Your machine · API | Free plan | Available | -- | |
| Cloud · Your machine | $19/mo | Available | 3.5 |
OpenPilot itself costs nothing. It is MIT licensed, needs no account and has no paid plan. What you pay for is the model: tokens are billed by whichever OpenAI-compatible provider you connect, and a local model costs only your hardware. Web search is optional and needs a TinyFish key.
Yes. The source is public on GitHub under the MIT licence, so you can fork, audit and package it. The project sits under an individual account rather than a company, and the FAQ says it was built from scratch rather than forked from Cline.
Yes. Anything that speaks the OpenAI API works, including Ollama, LM Studio, OpenRouter and your own gateway. The FAQ says smaller local models handle scaffolding well, while longer refactors usually want a larger model.
Partly. The FAQ says the app runs locally with no OpenPilot servers, but the README and the source show anonymous usage reporting that is on by default and goes to a telemetry host run by the maker. The README says it carries app opens, feature use and token counts, not chats, files or keys, and it can be switched off in Settings.
The download page offers only Windows and says macOS is coming soon. The GitHub release page already lists Apple silicon and Intel disk images, so a Mac build exists, but the vendor does not yet present it as supported.
OpenPilot is an open-source desktop agent that you point at a folder and a model of your choice. You paste in a base URL, an API key and a model id for any OpenAI-compatible endpoint, such as a cloud API, OpenRouter or a local server, and the agent reads, writes and edits files, runs terminal commands and streams the output back into the chat. It is an agent rather than a chat window because it plans a task into tool calls and carries them out in a loop, stopping for approval on sensitive actions. GitHub shows the repository created on 27 July 2026 under the account ibmshaikh, whose profile gives the name Ibrahim and a location in Mumbai; the site names no company.
It is a local Electron app with no OpenPilot account. On 22 November 2026 the main branch source was downloaded from GitHub and its telemetry module read: usage reporting is enabled by default and posts to a host at telemetry.openpilot.in, so the FAQ's claim of no servers is not literally true, though the README says no chats, files or keys are sent. API keys sit in operating-system storage, and conversations stay on the machine. Web search and page fetch are optional and run through a TinyFish key, which Hokai also lists as TinyFish. MCP servers are configured in an mcp.json file and switched on per chat, skills are reusable playbooks, and an AGENTS.md file holds project preferences. Local models work through Ollama or LM Studio, though its FAQ says smaller local models suit scaffolding better than long refactors.
It fits developers who already pay for model access and want a chat-first desktop agent without a subscription. See the wider coding agents list.
It is a poor fit for teams that need a mature, widely used agent: the project is months old, with one open issue and no release since 12 August 2026. OpenCode is MIT licensed and its GitHub repository showed about 228,000 stars on 22 November 2026, Cline lives inside VS Code and JetBrains, and OpenHands can open pull requests without per-step approval. The call: try it if you want a free desktop agent on Windows and already hold an API key; choose OpenCode or Cline if you want editor integration or a bigger user base.
Autonomy, guardrails, integrations and pricing on this page are read from Ibrahim (ibmshaikh)'s own documentation. The verdict is Hokai's reading of them, not the vendor's claim.