Alter

A small, fast desktop AI companion that remembers you. Bring your own model, or use your Claude Code subscription.

Alter is a native desktop chat app built on Tauri, so it is a few MB and light on memory. It talks to any OpenAI-compatible endpoint or drives the local claude CLI, keeps long-term memory across conversations, works inside a folder you attach, runs saved prompts on a schedule, and ships with a browser extension that reuses your Alter connections on GitHub and Frappe Helpdesk without putting a single API key in the browser.

Highlights

Models

• Any OpenAI-compatible endpoint. Presets for DeepSeek, Moonshot/Kimi, Gemini and OpenRouter are quick-fills, not an allowlist. Requests go through the Rust backend, so any host works.
• Claude Code as a backend: add one connection and use your claude subscription with no key. Each chat keeps a warm claude process, so follow-ups are instant and chats never interrupt each other. Pick Opus, Sonnet or Haiku and an effort level per chat.
• Per-chat connection, model and effort. Automatic fallback to another saved connection when the active one is down.
• Context meter in the composer: tokens used against the model's window, plus session cost.

Chat

• Streaming markdown with syntax-highlighted code, KaTeX math, tables and collapsible details, in light and dark themes.
• Esc interrupts the current turn at its next step and keeps everything done so far in the chat. Typing while a turn runs queues the message and delivers it at the next step, the same way Claude Code does.
• Attach images (persisted to disk), PDFs and text files. Drag and drop or paste.
• Branch a chat, regenerate, edit messages, export to Markdown, search across chats, pin favourites.
• Artifacts: HTML and SVG the assistant produces open in a sandboxed side panel.

Agent

• Modes: Auto (acts freely, writes confirm), Ask first, Plan (no actions) and Chat only.
• File tools on an attached folder: tree, grep, read, write. Web search and page reading. Every step shows in the conversation as it happens.
• Tool outputs are carried into later turns, so "now use that to…" works on every provider.
• Skills: save reusable instructions in the app, or use your Claude Code skills from ~/.claude/skills and ~/.claude/commands. Both appear in the slash menu, so /fix-issue 123 works inside Alter.

Memory

• Alter asks the model to tag durable facts as it replies, extracts them, and carries them into every future conversation. Review or forget any of them in Settings.
• The same facts are appended to ~/.claude/CLAUDE.md, so Claude Code learns them too, and a new Claude Code chat is seeded with what you told Alter.

Automation

• Routines: a prompt plus a schedule (interval, daily, weekly). Each run lands in its own chat. Runs execute in the Rust backend, so they fire with the window closed, and optionally as a login service.
• Projects group chats under a folder and standing instructions.
• ⌘K command palette, ⌘⇧Space global hotkey, menu-bar tray, launch at login.

Browser extension

• A token-gated bridge on 127.0.0.1 lets the companion extension use your Alter connections. No keys in the browser.
• GitHub: a maintainer-lens PR review with a plain-English verdict, verify-on-bench, and a draft comment in your own voice.
• Frappe Helpdesk: summarize, diagnose and draft replies on a ticket, then hand off to a full Alter chat or prepare a fix on a local branch for your review.
• Grammar fix on any editable field, page summaries from the popup.
• Details and setup in extension/.

Try the prototype

A clickable walkthrough of the whole app, built with frappe-ui: sketch.netchamp.dev/u/iamejaaz/alter. Send a message, press Esc mid-turn, open the slash menu, and click through Routines, Skills, Settings and the browser extension panels. Demo data only.

Quick start

Prerequisites: Rust (stable), Node 18+, and on macOS the Xcode Command Line Tools (xcode-select --install).

npm install
npm run dev

On first launch open Settings, pick a provider preset and paste a key, or add a Claude Code connection (needs the claude CLI on your PATH, logged in). Test connection checks the endpoint and lists its models.

Build a bundle with npm run tauri build. The .app and .dmg land in src-tauri/target/release/bundle/.

Settings worth knowing

Setting	What it does
Connections	One entry per provider or Claude Code. Keys stay on your device.
Agent working folder	Where browser-triggered agents run (PR review, support, prepare fix). Usually your bench or repo.
Frappe credentials	Site URL and API key/secret so the support agent's fr reads from the environment instead of the keychain.
Repro benches	Per-version bench folders the support agent can reproduce bugs on.
Memory	Everything Alter remembers, with delete.
Browser bridge	The pairing token for the extension.
Theme, launch at login	Under Settings.

Slash commands

/new, /clear, /usage, /compact, /auto, /ask, /plan, /chat, /folder, /attach, /routine <description>, /routines, /settings, plus every saved skill and every Claude Code skill as /<name> <args>.

Always-on routines

Routines run whenever Alter is open, including in the tray. To keep them running after you quit, install the background service after building:

npm run tauri build
cp -R src-tauri/target/release/bundle/macos/Alter.app /Applications/
./scripts/install-service.sh

Remove it with ./scripts/uninstall-service.sh.

Security notes

• OpenAI-compatible connections get file tools only: read, search and write inside the attached folder, no shell.
• Claude Code connections are full agentic sessions and can run commands. Auto mode runs them without prompts. Use Ask or Plan mode when you want to approve each step.
• Browser-triggered agents run under an allowlist: read-only fr and gh, read-only git (push, config, reset, commit and the like are denied), plus a scoped bench reproduction helper. This is a gate, not a sandbox.
• Everything (chats, memory, settings, attachments) is stored locally. Nothing leaves your machine except requests to the providers you configured.

Stack

Tauri 2 · Rust · React 18 · TypeScript · Tailwind CSS · Vite

License

MIT