Rowboat is an open-source desktop AI coworker, licensed Apache-2.0 and written in TypeScript, that maintains a memory of your work as a local Markdown knowledge graph and provides built-in surfaces — email, notes, browser, code mode, meeting notes — for acting on that memory.

What it is

Rowboat is a desktop application for Mac, Windows and Linux that sits in the Infrastructure & Operations / Orchestration & Scheduling category. It indexes email, meetings, Slack and assistant conversations into an Obsidian-style backlinked knowledge graph, then uses that graph to get work done on the machine where it runs. Alongside the graph it ships work surfaces for collaborating with AI: an email client, notes, a browser, code mode, a meeting note-taker, and separate workspaces for different projects. Models are bring-your-own — local via Ollama or LM Studio, or hosted providers with your own API key.

The concrete problem it addresses is cold-start context. Most AI tools reconstruct context on demand by searching transcripts or documents, so every session begins from nothing and relationships between pieces of work stay implicit. Rowboat keeps long-lived knowledge instead: context accumulates over time, relationships are explicit and inspectable, notes remain editable by the user rather than hidden inside a model, and everything lives on your machine as plain Markdown. It replaces per-tool, per-session context rebuilding with a vault that persists and compounds.

Key capabilities

• Brain: indexes email, meetings, Slack and assistant conversations into a backlinked knowledge graph stored as plain Markdown on the local machine.
• Built-in email client: sorts mail into important and everything else, and automatically drafts responses for important email using accumulated work context.
• Background agents: run on events such as a new email or on a schedule such as every day at 8am, with access to tools, web search, the browser, and code written through Claude Code or Codex.
• Meeting notes: a local note-taker taps mic and speaker, produces a live transcript, summarizes into a Markdown file, and updates the knowledge graph.
• Code mode: spins up parallel coding agents using Claude Code or Codex and drives them with the relevant work context.
• Isolated browser: a built-in browser kept separate from the main browser, so you log in only to the accounts the assistant may access.
• Extensibility through MCP: connect any MCP server or Composio tools, for example Parallel Search MCP, which exposes web_search and web_fetch without an account or API key under rate limits, added through Settings → MCP Servers.

Who uses it and how

• Individuals on Mac, Windows or Linux who install the desktop build and keep their knowledge vault locally.
• Knowledge workers with heavy email and meeting load who connect Gmail, Calendar and Drive by following the Google setup guide.
• Developers who run parallel coding agents through code mode with Claude Code or Codex.
• Privacy-conscious users who keep data in a local Markdown vault and run local models through Ollama or LM Studio.
• Users building custom work surfaces inside Rowboat that reuse existing tools and integrations and can be shared with other people.

Getting started

Download the latest build for Mac, Windows or Linux from https://www.rowboatlabs.com/downloads, or take any release file from https://github.com/rowboatlabs/rowboat/releases/latest. Optional keys go in ~/.rowboat/config/deepgram.json for voice input, elevenlabs.json for voice output, exa-search.json for Exa research search, and composio.json for Composio tools, all in the format { "apiKey": " " }.

How it compares

The registry provides no list of paid products that Rowboat replaces, so it is best understood as an orchestration layer rather than a substitute for the tools it names. It does not replace Claude Code or Codex — it drives them — and it sits above whichever model runtime you choose, whether Ollama or LM Studio locally or a hosted provider by API key, under the Apache-2.0 licence.

When to use it — and when not to

A user must operate a desktop install with its own configuration surface: a local Markdown vault, either a local model runtime or hosted provider credentials, Google OAuth consent for Gmail, Calendar and Drive, and optional keys for Deepgram, ElevenLabs, Exa and Composio. It is a poor fit for teams that need a hosted, multi-user cloud service, or for environments where work data may not live on an individual's machine. The provided README excerpt also ends mid-sentence in the Parallel Search MCP setup, and no server or hosted deployment path is documented, so organisation-wide rollout and audit expectations are unclear.

{
  "apiKey": "<key>"
}

{
  "mcpServers": {
    "parallel": {
      "url": "https://search.parallel.ai/mcp"
    }
  }
}