Skip to content

About

AI writing, rewriting, summarization, and translation directly in Thunderbird, using your chosen provider or local model.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

3 stars

Watchers

0 watching

Forks

ThunderClaw character

ThunderClaw

AI writing and reading assistance for Thunderbird, using your chosen provider or local model.
Your inbox, your decisions, your control.

CI Apache 2.0 license Thunderbird 128+

Features · Safety · Installation · Development · Contributing

Install ThunderClaw from Thunderbird Add-ons. The standalone extension connects directly to your chosen provider or local model.

AI assistance without handing over your inbox

ThunderClaw brings useful writing and reading tools into desktop Thunderbird while keeping Thunderbird—and you—in charge. A model can propose an edit, summary, or translation. It cannot send mail, silently rewrite a message, or change recipients, attachments, or headers.

Every compose change follows the same deliberate flow:

Generate → Preview → Apply → Undo

The current Thunderbird content remains authoritative throughout. Results are validated in the background and at the Thunderbird editing boundary, and stale results fail closed.

What it can do

In the composer In a received message
Improve, proofread, shorten, or change tone Create a separate plain-text summary card
Translate or summarize selected draft text Reversibly translate visible message text
Follow a custom writing instruction Preserve the message's original HTML structure
Preview before applying and undo afterward Leave the original message untouched

Thunderbird 128 and newer support plain selected-text transformations. Qualified Thunderbird 153 and newer selections can also preserve or produce narrowly typed rich content: paragraphs, flat lists, and bold, italic, or underlined spans. Unsupported shapes are rejected rather than guessed at.

See the product contract for exact capability and eligibility rules.

Your inference connection

ThunderClaw connects directly to OpenAI, OpenRouter, DeepSeek, Fireworks AI, Ollama, LM Studio, or a custom OpenAI-compatible endpoint. Choose the endpoint and model in settings, provide a key if needed, grant host access, save, and run the synthetic connection test. Keys are remembered locally across restarts. They are not synchronized or an encrypted vault.

Compose requests send the selected text and authored draft context, without recipient addresses or extracted quoted history. Reading requests send eligible visible message segments. Local state fingerprints still detect changes to recipients, headers, attachments and the captured text before Apply. Read the privacy policy for details.

A loopback endpoint can itself forward to hosted models. OpenRouter may route to an upstream provider. ThunderClaw never silently changes your configured endpoint or model. The project operates no inference or analytics service.

How it stays safe

  • Thunderbird owns DOM changes, validation, rollback, and Undo. Models return bounded typed data; model HTML is never inserted.
  • Sending remains manual. Models cannot change recipients, headers, or attachments.
  • Request/run identity, generations, target/context hashes, connection epochs, output limits and exact segment mapping reject stale or misplaced changes.
  • The background owns credentials, prompts and transport. Settings and feature messages are restricted to their trusted extension pages.
  • HTTPS is required remotely; explicit loopback HTTP is supported. Redirects and automatic endpoint/model fallback are rejected.

See architecture and security and privacy.

Installation

Install from Thunderbird Add-ons, then open ThunderClaw settings. Choose a provider, load a model where available or enter its exact ID, save your key and connection, grant host access, and click Test connection. Your key stays saved across restarts.

For local development, build the XPI from this checkout and install it through Thunderbird’s Add-ons and Themes → gear menu → Install Add-on From File. See installation and connection setup for provider presets, local servers, troubleshooting, and upgrade guidance. Installing a newer XPI updates the existing add-on; uninstalling first is unnecessary.

The response limit defaults to 16,384 tokens. Thinking defaults to Off for DeepSeek and Model default elsewhere. Both are configurable; model support varies. See response controls.

Development

ThunderClaw uses Node.js through mise:

mise exec -- npm ci
mise exec -- npm run check

The main source and verification paths are:

packages/thunderbird-extension/   Extension source, settings, and icons
test/                             Deterministic unit and boundary tests
e2e/thunderbird/                   Real Thunderbird harnesses and fixtures
fixtures/                         Conformance fixtures
scripts/                          Build, packaging, and qualification tools
docs/                             Product, engineering, security, and release docs

Start with the development guide, testing guide, and contribution guide. The roadmap tracks unfinished qualification and hardening. Bug reports, documentation improvements, tests, and focused changes are welcome.

Documentation

Community and license

Please read CONTRIBUTING.md before opening a pull request and use SECURITY.md for private vulnerability reports. Community participation is governed by the Code of Conduct.

Code, documentation, and project-owned assets are licensed under the Apache License 2.0. The ThunderClaw name, logo, and character are also covered by the trademark policy.


Want safer AI assistance in Thunderbird?
Star ThunderClaw to follow the journey—and help more Thunderbird users find it. ⚡

About

AI writing, rewriting, summarization, and translation directly in Thunderbird, using your chosen provider or local model.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

3 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages