Local-first agentic work
What Tidebreak is
A local-first AI agent desktop app that runs on your machine with the models and tools you choose.
Tidebreak is a desktop app for handing real work to an AI agent, and it has two surfaces. In a chat, you attach the files the work depends on and the agent reads them, runs code, searches the web, and writes the result out as a file you can save. In code mode, you point it at a git repository and supervise a coding agent working in an isolated worktree.
It runs on your machine. Your conversations, your documents, and the files the agent produces live in a local database and a local folder. Your provider API keys live in your operating system's credential store. Nothing is routed through a service the project operates — the only outbound traffic is the requests you configure: your model provider, your search provider, and whatever the agent explicitly fetches.
Tidebreak is pre-1.0 and under active development. Interfaces, storage formats, and behavior change frequently. Expect rough edges.
What you can do with it
- Work over your own files. Attach documents to a chat from the composer, or connect a folder on your machine that the agent can list and read.
- Run code in a sandbox. The agent writes and executes scripts to analyze spreadsheets, render PDFs, build charts, and generate documents. Commands run inside the code-execution backend you choose. Network access is a separate per-chat policy that starts open and can be narrowed to package registries, exact hosts, or fully offline.
- Produce files, not just answers. Anything the agent saves to its
output/directory becomes a versioned output you can preview and export with a native save dialog. - Build reusable local apps. Ask for a triage view or other small tool once, then reopen it from the Apps library. Its access is pinned to the exact folders and connected-app operations you approve.
- Connect outside tools. Add MCP servers for chat tools, or a REST API from an OpenAPI document for local apps to call through Tidebreak's governed executor.
- Use the model routes you prefer. Anthropic, OpenAI, Google, xAI, Fireworks, Together, OpenRouter, a local Ollama daemon, and any OpenAI-compatible endpoint. You can switch models mid-conversation without losing the thread.
- Drive coding agents over a repository. Code mode runs Claude Code, Codex CLI, opencode, or Grok CLI in its own git worktree per workspace, with per-turn diffs, a review step, a terminal, and a pull-request flow.
- Decide how much rope to give it. Every chat and every coding session has a permission mode, from read-only planning up to full autonomy.
Who it's for
People who want an agent that can touch real data and produce real artifacts, and who would rather that happen on their own machine than in someone else's cloud. You can sign in with a ChatGPT subscription, provide an API key, or use a local Ollama model.
What it is not
It is not a hosted service, and there is no Tidebreak account to create. It is not primarily a repository-bound coding assistant — the unit of work is a chat with files, folders, tools, and outputs. And it is not finished: the product is still pre-1.0 and its local data model can change between releases.