AI You Control: Choose your models. Own your data. Eliminate vendor lock-in.
Thunderbolt is an open-source, cross-platform AI client that can be deployed on-prem anywhere.
- 🌐 Available on all major desktop and mobile platforms: web, iOS, Android, Mac, Linux, and Windows.
- 🧠 Compatible with frontier, local, and on-prem models.
- 🙋 Enterprise features, support, and FDEs available.
make doctor # verify your tools — prints exact install commands for anything missing
make setup # install frontend + backend dependencies, wire up agent symlinks
make up # start Postgres + PowerSync in Docker
make run # start the backend (:8000) and frontend (:1420)For self-hosting with Docker Compose, Kubernetes, or Pulumi on AWS, see docs/self-hosting/. For full dev-environment details, see docs/internals/development/quick-start.md.
Found a bug? Have an idea?
- We're actively working on our docs, community, and roadmap. For now, the best way to get in touch is to File an issue.
We welcome contributions from everyone.
- Getting set up: the devel 8000 opment guide gets the backend, sync service, and frontend running locally.
- Conventions: AGENTS.md documents the code style and architecture invariants this repository is held to — TypeScript and React rules,
useEffectdiscipline, localization, responsive sizing, the app-version gate, reconciled defaults. It is written for coding agents, but it is the same reference human reviewers use. - Further development, tooling, and architecture docs are listed under Documentation below.
- Make sure to check out the Mozilla Community Participation Guidelines.
Everything under docs/ is also published at thunderbolt.io/docs.
- Introduction — what Thunderbolt is, and who it is for today
- FAQ — funding, the relationship to Thunderbird, model support, data handling
- Customize — the extension points: agents, models, skills, projects, widgets, MCP servers, auth providers
- CLI —
thunderbolt, the single-binary terminal coding agent, and the ACP/MCP bridge that connects it to the app - Telemetry — what is collected, and how to turn it off
Features:
- Projects — a workspace whose instructions every chat inside it inherits
- Skills — reusable instruction bundles invoked with
/slugor loaded by the model - Widgets — interactive components the model embeds in a response
- HTML Artifacts — model-authored pages rendered in a sandboxed iframe
- Attachments — files on a chat turn, and why the bytes are never stored server-side
- Voice Mode — spoken conversation through the same send path as typing
- Search and the Command Palette —
Cmd/Ctrl+Kover chats, messages and settings - MCP Connections — adding Model Context Protocol servers
- ACP Agents — handing a thread to an external coding agent over WebSocket or iroh
- WebView — opening links in the app's side panel (desktop and mobile)
- Multi-Device Sync and End-to-End Encryption — local-first SQLite, opt-in sync, opt-in E2EE
- Export Format — the JSON snapshot behind Settings → Preferences → Export My Data
- Overview — the three deployment targets and the stack they share
- Configuration — every backend environment variable
- Docker Compose — single-host stack, for demos and evaluation
- Kubernetes — manifests and ConfigMaps synthesized from
deploy/config/ - Pulumi (AWS) — ECS Fargate or EKS via infrastructure-as-code
- Deployment assets — the Dockerfiles, realm and sync-rule configs all three targets build on
- Authentication — the four flows that mint a session, plus OIDC, SAML and personal access tokens
- Rate limiting — the Postgres-backed limiter in front of the routes that cost money
- Self-hosting the iroh relay — for the peer-to-peer CLI↔app bridge
- Quick Start — prerequisites, bootstrap, and the local service layout
- AGENTS.md — code style and architecture invariants (see Contributing above)
- Frontend Structure — where a new file goes in
src/, and theui/conventions - Error Handling — optimistic code, and the few places that legitimately catch
- Testing and Backend Testing —
bun testscopes, PGlite, and what not to run at the repo root - Mobile Setup — iOS and Android Tauri dev
- Integrations — adding a third-party account the model can act on
- CI and Preview Environments — the workflows a pull request starts
- AI Code Review — the automated review every non-draft PR receives
- Storybook, Vite Bundle Analyzer, Local CDN for app updates, Tauri Signing Keys
- Release Process — cutting and publishing a release
Start with the architecture map — the components, how they talk, and where each piece of state lives. Then, by area:
- Client — App initialization · Chat runtime · System prompt, tools and citations · Content view · Data access layer · Auth and session · Client data migrations · Settings and preferences · Reconciled defaults · Tauri shell · In-browser agent harness · The
shared/module - Backend — API surface · Universal proxy · Managed inference · Sign-in and the waitlist · Debug transcripts · Backend service
- Sync and data — PowerSync, account and devices · Sync middleware · Upload authorization · Composite primary keys and default data · Delete account and revoke device
Please read our Code of Conduct. All participants in the Thunderbolt community agree to follow these guidelines and Mozilla's Community Participation Guidelines.
If you discover a security vulnerability, please report it responsibly via our vulnerability reporting form. Please do not file public GitHub issues for security vulnerabilities.
Thunderbolt is licensed under the Mozilla Public License 2.0.