Using Thunderbolt
Thunderbolt is an open-source AI workspace. Chat with its built-in assistant, or bring your own agent, like Claude Code or Codex.
Chats
A conversation is titled from your first message. A message you have already sent cannot be edited, and a reply that succeeded cannot be regenerated; ask a follow-up instead.
Chat covers attachments, artifacts, citations, and what the app shows when a reply fails.
Models and agents
Each conversation keeps its own model and agent, picked from the composer and the chat header.
Settings → Models and the composer’s model picker cover the built-in Thunderbolt agent only. Every other agent, whether your deployment serves it or you added it yourself, chooses its own model upstream. The picker is hidden while one is selected, and whether you get any say in that choice is up to the agent rather than Thunderbolt.
Adding models and providers is covered in Customize.
Projects
A project is a set of instructions that every chat inside it inherits, plus a view of the chats that belong to it. Deleting a project keeps its chats.
The assistant can keyword-search the project’s other conversations when it needs earlier context, so wording that differs from the original chat can miss. Projects do not hold an uploaded document set; files are attached per message, in the chat that needs them. See Projects.
Skills
A skill is a named block of instructions. Type / or @ in the composer to pick one for a single message, or let the assistant load one on its own when the request calls for it. We ship a set of skills, and you can add your own under Settings → Skills. Pin the ones you use most, up to ten, and they appear as chips above the composer when you start a chat. See Skills.
Connections
Settings → Connections is where the assistant gets tools beyond the ones built in.
| Connection | What it adds |
|---|---|
| Thunderbolt | Web search and page fetching. On by default; both need a search provider key on the server |
| Google, Microsoft | Google: mail and calendar, plus draft creation. Microsoft: Outlook mail and OneDrive files, read-only |
| An MCP server you add | Whatever tools that server publishes, such as searching a wiki, querying a database, or filing a ticket |
MCP is the Model Context Protocol, an open standard for publishing tools to an AI assistant. You add a server by URL, and a test button lists the tools it offers before you save it. Google and Microsoft also need whoever runs your deployment to have configured sign-in credentials for that provider.
External agents are separate and are added under Settings → Agents. Connections covers both.
Voice
When the composer is empty, the send button becomes a voice button. Thunderbolt transcribes what you say, sends it as an ordinary message, and speaks the reply. Talking over the reply interrupts it. Only your finished utterance is sent; there is no live microphone stream, and audio is never stored. See Voice.
Search
Cmd+K on macOS or Ctrl+K elsewhere searches chats, messages, models, skills, agents, connections, devices, projects and tasks, and runs commands such as New chat.
Search runs against an index on the device, so it works offline and no query leaves the machine. It matches words, not meaning: a question phrased differently from the text you are looking for will miss. See Search.
Where settings live
| Page | What you set there |
|---|---|
| Agents | The agents available in the agent picker |
| Skills | Create, edit, and enable skills. Pinning is done from the composer’s skills bar |
| Connections | MCP servers, external services, and the tools each one exposes |
| Models | Providers, API keys, and which models appear in the picker |
| Voice | The speech provider, when the voice preview is turned on |
| Preferences | Appearance, your name, language and units, privacy, network proxy, sync, export, deletion |
| Devices | The devices signed into the account, and revoking one |
Your preferred name, location, language, units, and usage-data preference follow the account. Theme, haptics, link behaviour, and voice provider stay on this device, along with model API keys, MCP servers and their credentials, and whether this device syncs at all.
Attached files stay on the device that added them too: the message syncs, the file does not. Apps and Sync describes what replicates between devices and how a new device is approved.
Preview features
Tasks is a to-do list the assistant can also read and write. It is off by default and is enabled under Preferences → Preview Features, which adds a Tasks entry to the sidebar. Where a deployment has usage analytics configured, Tasks also requires anonymous usage data to be on, and turning that off turns Tasks off.
The custom voice provider and cross-device sync are in preview as well, and the end-to-end encryption that protects synced data has not had a cryptography audit. Everything else described here is on by default.