Apache-2.0 · Local-first · Built in Rust
Proven on three years of private equity diligence and rebuilt in Rust. Grant a folder, pick a model, and walk away with a spreadsheet, deck, or memo you can put your name on.
Sign in with ChatGPT. The OAuth handshake runs on your own machine, and you use the subscription you already pay for. No API key to create.
Anthropic, OpenAI, Google, xAI, Fireworks, Together, or OpenRouter. The key goes in the OS credential store, not in a config file you might commit.
Point the OpenAI-compatible slot at Ollama and run open-weight models on your own hardware, so even prompts stay on the machine.
Work mode and code mode ride the same engine. Swap any layer under either one.
Start on the ChatGPT plan you already pay for.
Anthropic, OpenAI, Gemini, xAI, Fireworks, Together, OpenRouter, or a local Ollama model. Switch mid-chat. Cheap model for the grind, strong model for the hard step. In code mode each engine keeps its own sign-in, and you set the model, reasoning effort, and fast mode from the composer.
Same exec tool. Different backends.
The native sandbox is the macOS default. On Windows and Linux, pick Docker, E2B, or Daytona. Per-chat network policy: offline, package installs, custom hosts, or open. Credentials never enter the sandbox. A code workspace can also run on a Model Gateway machine that clones, commits, and pushes as you.
A folder is available only after you pick it.
Grant each capability on its own. Disconnect a folder from one chat, or forget it everywhere. Overwriting a connected file always asks, in every permission mode. In code mode each workspace gets its own git worktree under a root you can see and move, so no agent touches your working copy.
Computer use (screen and app control) is partial, and macOS-only.
Look things up without leaving the machine.
Anthropic and OpenAI can search through the provider you already configured. Or bring Exa, Tavily, Brave, Firecrawl, or a SearXNG you host. On macOS the agent can drive an in-app browser with its own profile and hand it back when it needs you. MCP servers and REST APIs registered from an OpenAPI document reach the rest. A managed OAuth catalog is deliberately not the plan.
Finished files and reviewed diffs.
A spreadsheet, a deck, a memo, a chart, a small app, or a pull request. Built-in skills cover Word, PDF, PowerPoint, spreadsheets, and charts. Code mode drives Claude Code, Codex CLI, opencode, and Grok CLI in isolated worktrees, opens the pull request, and keeps repairing it until the checks are green. Nothing reaches your branch until you have read the change.
The engine runs on your machine. The defaults stay boring on purpose.
The same four permission modes govern both surfaces, so a coding agent asks before it acts on the terms a document agent does. Every MCP tool asks. An agent can open, repair, and update a pull request, but the merge button only answers to you. Rust, because the loop is fast enough to feel like a native tool.
Features
Switch models, grant only what a task needs, and read every result before you keep it, whether that is a finished file or a change on a branch.
Preview any version. Restore one without deleting what came after.
v6 · You
Restored from v2
v2 · Agent
23 min ago
v1 · You
31 min ago
One task. One journal. Any provider.
Research complete
OpenAI · 12 journal entries
Drafting the final brief
Claude · context carried forward
Access starts when you grant it and ends here.
Q3 diligence
Connected folder
Your files stay yours in every permission mode.
Replace board_deck.pptx?
You edited this file after the agent’s last version.
Read what the agent changed before any of it reaches your branch.
<section class="hero">
− <h1>The research engine</h1>
+ <h1>One engine. Two modes.</h1>
</section>
Each one works in its own git worktree, so parallel work never collides.
Watch and fix survives a restart, repairs each new failure, and stops when the PR is ready for you.
#2248 · hero-lab
Claude Code · watching since 9:14
Checks failed, conflicts, changes requested, ready to merge. Each trigger fires once per commit.
The world we're building toward
Tidebreak is the open surface. The harder work is making every model, company app, and production system usable by agents without giving up control. That is the system we are building around it.
Models + apps
Bring Anthropic, OpenAI, Google, xAI, local models, and MCP-connected applications behind shared identity, grants, routing, usage, and spend controls. Tidebreak already signs in to it for hosted models and machines.
Infrastructure
Give agents policy-bound access to Kubernetes, PostgreSQL, cloud CLIs, SSH, and SaaS through short-lived credentials, explicit approvals, and a durable audit trail.
Code
Code mode already drives four engines in isolated worktrees, reviews every turn as a diff, watches a pull request until it is green, and merges it from the app. What is ahead: sessions you hand to a hosted machine from Slack or your phone, a scheduler, and tests an agent is trusted to act on.
Design partners
We’re looking for teams to shape how agents work across models, software, and infrastructure.
For three years we built enterprise-grade research agents for some of the most high-fidelity work there is: private equity diligence, the kind of file someone stakes their career on.
Now we're shifting our focus to secure, compliant infrastructure for the agentic AI era. Tidebreak is the open client at the edge of that system.
macOS, Windows, and Linux. Your ChatGPT plan, a key, or Ollama. No Tidebreak account.
