Desktop AI workspace · Windows, Linux

The AI workspace that stays on your machine.

A chat client, a coding agent and connectors to the tools you already use, in one desktop app. Bring your own API keys. Nothing leaves your computer except the requests you choose to make.

Coming soon · Free and source-available · No account

Illustration of a Gantry session: the user asks the agent to find out why the auth tests fail. The agent reads the test and the module, edits two files, runs the tests, and asks permission before opening a pull request on GitHub.

Runs locally

One desktop app. Your chats live in a database file on your disk.

Your own keys

Add keys from Anthropic, OpenAI, Google, xAI or OpenRouter. Pay them, not us.

Source-available

Read the code, build it yourself, file an issue. FSL, Apache 2.0 after two years.

What agents can do

Watch it work. Every action is a row in the feed, and every risky one waits for your yes.

Edits files, runs commands, shows everything

The agent works inside the folders you attach: targeted edits arrive as diffs you can revert, commands stream their output, and each tool call sits in the activity feed with its input and its result. Nothing happens off screen.

See the agent
ActivityAuto-edit
  1. filesystem read_file src/lib/auth.rs 212 lines
  2. code-editor str_replace src/lib/auth.rs +6 −2
  3. shell cargo test --package api 31 passed
  4. postgresql query select count(*) from sessions 1 row
  5. github create_pull_request allow?
An illustrative session: file read, a targeted edit shown as a diff, tests run, a database query, and a pull request waiting for permission.

Asks first, on your terms

Four permission modes per chat, from approve-everything to hands-off with a guard model reviewing each risky call. Reads run freely; writes outside your folders, commands and anything destructive follow the mode you chose.

How permissions work

github wants to create_pull_request

Fix token expiry comparison · api ← fix/token-expiry · write_external

Allow onceAlways for this chatDeny
Auto-edit mode lets reads and edits run; an external write such as opening a pull request produces a prompt with allow once, always and deny.

Connects to the tools you already use

Built-in connectors for files, editor, shell and web, and MCP servers for GitHub, Linear, Notion, Sentry, Figma, Supabase, Stripe and dozens more. The same permission model applies to all of them, and nothing installs itself.

Browse the connectors

What Gantry is

One workspace, five things. A chat client, a coding agent and connectors, kept together by projects and your own keys.

  • Chat with any model

    Claude, GPT, Gemini and Grok, or anything on OpenRouter and Ollama, from one window. Switch per chat. Keys go to the provider, never through us.

    • Anthropic
    • OpenAI
    • Google Gemini
    • xAI
    • OpenRouter
    • Ollama
    • Mistral
  • Projects

    Group chats with the folders, files and instructions they share. A project remembers its context so each chat starts informed.

  • Artifacts

    Documents, code, diagrams and small interactive apps the model builds live beside the chat, versioned, and run in a sandbox that cannot reach your files.

  • Skills and memory

    Reusable instructions and remembered facts, both plain text you can read and edit. The assistant proposes; you confirm. Nothing is remembered silently.

  • Windows and Linux

    One app, built with Rust and Tauri, with macOS to follow. Your data is a SQLite file next to your attachments; back it up, move it, delete it.

Local by construction

No account, no server, no telemetry. The app talks to the providers and services you configure, and to nothing else.

  1. FIG 0.1

    Keys stay in your keychain

    API keys and tokens are encrypted with a key that lives in your OS keychain, Credential Manager or Secret Service. They never reach the interface layer, and they never reach us.

  2. FIG 0.2

    Your data is a file you own

    Chats, projects, memory and artifacts live in a SQLite database and a folder on your disk. Full-text search included. Copy it, back it up, delete it.

  3. FIG 0.3

    Nothing installs itself

    The connector catalog ships with the app, and every connector waits for an explicit Install. The agent can suggest one; only you can add it.

In the open

Source-available, built in public. Read the code, build it yourself, tell us what is missing.

Gantry is licensed under the Functional Source License: you can read, run, modify and share it, and each release becomes Apache 2.0 two years on. Development happens in one repository with the plan documents beside the code.

oljodev/gantry A desktop AI workspace: chat client, coding agent and connectors. Rust, Tauri, React. ★ —FSL-1.1-ALv2No release yet

Get Gantry

Free. No account. Add a key and start.

Requirements, checksums and building from source on the download page.