Everything an AI workspace needs, on your desktop.
Five chapters: the chat, the agent, permissions, connectors, and what stays local. Each one links to the docs that go deeper.
Chat
Any model, one window.
Talk to Claude, GPT, Gemini and Grok models, or anything on OpenRouter or a local Ollama, from one place. Keys go straight to the provider. Projects group chats with the folders, files and instructions they share, and artifacts, the documents and small apps a model builds, live beside the conversation, versioned, in a sandbox that cannot reach your files.
Summarise what changed in the auth module this week and draft release notes.
Three changes landed: the token expiry fix (#412), rate limiting on /auth/login (#415) and the session cleanup job (#418). Draft notes are in the artifact on the right; I kept the tone of the 0.3 notes.
Release notes 0.4Markdownv2
Agent
Edits you can read, commands you can see.
Attach a folder and the agent can read, search and change what is in it. Edits are targeted replacements shown as diffs, revertible from the feed. Commands run in your shell with their output streaming in. Every tool call is a row in the activity feed with its input and its result, so a session is something you can scroll back through, not a summary you have to trust.
- if token.expires_at < now_ms {+ if token.expires_at * 1000 < now_ms {
$ cargo test -p api running 31 tests … ok
Permissions
Four modes, chosen per chat.
Manual asks for everything. Auto-edit lets reads and edits inside your folders run, and asks for commands, external writes and anything destructive. Plan reads and searches but changes nothing. Auto runs everything, and with Guard on, a fast model from your provider reviews each risky call against the task and blocks what does not fit. The mode is a setting on the chat, so a scratch repository can run hands-off while a production database asks every time.
Manualask for everythingAuto-editreads and edits runPlanread onlyAutoguard on
Guard reviewed shell · rm -rf build/
Allowed: the path is inside the project and the command matches the task. Anything outside the folders you added would have been blocked.
Connectors
Built in, or any MCP server.
Four connectors are part of Gantry: the filesystem, the code editor, the shell and web fetch. Everything else is a Model Context Protocol server, run as a local process or reached over HTTPS, with a curated catalog that ships with each release — every entry in it checked against the service’s own server, not written from memory. Signing in happens in your browser and the token stays on your machine. Nothing installs itself: the agent can suggest a connector, only you can add it.
There is no account and no Gantry server. Chats, projects, memory and artifacts live in a SQLite database and a folder on your disk, with full-text search. API keys and tokens are encrypted with a master key kept in your OS keychain, Credential Manager or Secret Service, and never reach the interface layer. Back the folder up, move it, or delete it.
No account. No Gantry server. Delete the folder and it is gone.
kinetic-apiClaude Opus 5Auto-edit
Summarise what changed in the auth module this week and draft release notes.
Three changes landed: the token expiry fix (#412), rate limiting on /auth/login (#415) and the session cleanup job (#418). Draft notes are in the artifact on the right; I kept the tone of the 0.3 notes.
Release notes 0.4Markdownv2
ActivityAuto-edit
code-editorstr_replace src/lib/auth.rs+6 −2
shellcargo test -p api31 passed
code-editorstr_replace tests/auth_test.rs+9 −0
- if token.expires_at < now_ms {+ if token.expires_at * 1000 < now_ms {
$ cargo test -p api running 31 tests … ok
Manualask for everythingAuto-editreads and edits runPlanread onlyAutoguard on
Guard reviewed shell · rm -rf build/
Allowed: the path is inside the project and the command matches the task. Anything outside the folders you added would have been blocked.
InstalledCatalogAdd your own
FilesystemBuilt inon
GitHubMCP · signed inon
PostgreSQLMCP · localon
SlackMCP · not installedInstall
~/Library/Application Support/Gantry
gantry.db chats, projects, memory · SQLite
blobs/ attachments and artifacts
master key in the OS keychain, not on disk
No account. No Gantry server. Delete the folder and it is gone.
Risk tiers
Every tool call has a tier. The mode decides what each tier does.
Tier
Examples
Manual
Auto-edit
Plan
Auto, guard on
read
Read a file, list a folder, search issues
asks
runs
runs
runs
write
Edit a file in a folder you added
asks
runs
blocked
runs
write_external
Open a pull request, send a message, insert a row
asks
asks
blocked
guard decides
execute
Run a command or a script
asks
asks
blocked
guard decides
destructive
Delete files, drop tables, force-push
asks
asks
blocked
asks, always
app
Save a memory, create an artifact, propose a skill
asks
runs
runs
runs
Try it on your machine
Free, no account. Add a key from your provider and start.