Rowboat Review

Most AI assistants start from zero every conversation. Rowboat's pitch is the opposite: index email, meeting transcripts, Slack and assistant chats into a backlinked Markdown graph on your own disk, then let an email client, meeting note-taker and background agents work off that context. Code Mode spinning up Claude Code or Codex from your actual work context is a genuinely different move than a generic coding assistant.

Apache 2.0 licensing and local-first storage answer the data-portability question before the board asks it. Model-agnostic support for Ollama, LM Studio or your own API key is a real hedge against vendor lock-in.

Starter's $5-25/month range is cheap enough to pilot without a purchase order. I couldn't find docs or a changelog, so shipping cadence and integration depth are unverified. Pilot with a handful of knowledge workers for a quarter before touching Enterprise.

Storing the Brain as backlinked plain Markdown on local disk is the correct architectural call. If your knowledge graph lives in files you can grep, diff, and version, then Rowboat never becomes the single point of failure that hosted-only assistants do. MCP and Composio for integrations, plus Ollama/LM Studio for local inference, means you're not locked into one model vendor.

Code Mode driving Claude Code or Codex agents off your actual work context, with a reviewable diff, is a sane trust boundary for anything touching your codebase. But background agents that fire on schedule and drive a built-in browser autonomously are a different risk class — I'd want audit logs and rollback before letting one write code unattended against production.

Pricing scales from $5 to $200+/month plus a free self-hosted tier, so cost isn't the gate. The gate is whether your team wants a second runtime — desktop app, local models, MCP servers — sitting alongside your existing agent stack for the next three years.

$5/month Starter to $200/month Pro. That's not one price, that's two bands. 50 seats at Pro-band-high, $200 × 12 × 50 = $120K/year. At Starter-band-low, $5 × 12 × 50 = $3K/year. Forty-times spread on the same tier name. Procurement needs the exact number, not the range.

Self-hosted is free. Apache 2.0, run it on local hardware, no seat math at all. That's the real TCO lever — skip the cloud tier, add local compute cost instead. Ollama or LM Studio, bring your own hardware.

Enterprise is 'contact sales.' Shared memory and server-side background tasks gated behind a call — standard for that tier, still a procurement delay. 7-day trial on Starter, no free trial listed for Pro. No published auto-renewal terms, no stated overage rate. That's the invoice risk, not the sticker.

The Brain concept is the sell: backlinked plain Markdown on disk instead of a vendor-owned vector store, Apache 2.0 licensed so I can actually read what's indexing my email. Code Mode driving Claude Code or Codex off meeting context into a reviewable diff is the part I'd actually adopt first — it's a narrow, checkable output, unlike auto-drafted email replies which need trust built over weeks of edits.

Workflow fit depends on how much you already live in email/Slack/meetings versus a ticket queue or repo. Background agents on schedules and events are good primitives, but five work surfaces (email client, note-taker, browser, Code Mode, Apps) is a lot to context-switch between before habits form.

I couldn't find docs, API reference or a changelog, which matters more here than usual — MCP server config and local model setup (Ollama, LM Studio) are exactly the steps that go wrong first. Pricing scales oddly: $5 to $25 for Starter, $50 to $200 for Pro, both ranges rather than fixed numbers.

The pitch here is genuinely different: instead of a chat box that forgets you exist between sessions, Rowboat builds a Markdown knowledge graph — the Brain — out of your actual email, meeting transcripts and Slack history, stored on your own disk. That's a real answer to the thing that makes AI assistants feel disposable. Email triage, a meeting note-taker that listens locally instead of sending a bot into your call, and Code Mode kicking off Claude Code or Codex agents are concrete, specific features, not vague promises.

Day one is going to feel like homework, though. You're granting access to email, calendar, chat and meetings before the tool has proven anything to you, and the value only shows up once the graph has enough material to be useful — that's a week-three payoff, not a ten-minute one.

Pricing starts at $5/month Starter, jumps to $50 Pro, with no listed free trial on the paid tiers despite a free self-hosted option. Being Apache 2.0 and model-agnostic (Ollama, LM Studio, your own key) is the real trust move here.

Local-first, Markdown-on-disk, Apache 2.0. That's a real answer to the privacy question every AI coworker pitch dodges. Bring your own Ollama model, keep everything on your machine — fine, I believe that part, it's a specific architectural claim, not marketing fog.

But the pricing confuses me. Free and self-hosted exists, yet Starter runs $5-25/month and Pro jumps to $50-200 for 'effectively unlimited usage' — vague band pricing is the kind of superlative that ages poorly when you're trying to budget. No docs page or changelog I could find, but there is a blog. Could be there and just not surfaced. Could also mean a two-person team shipping fast and not writing it down yet.

Code Mode driving Claude Code or Codex is a nice differentiator versus a plain chat assistant. Exit is clean on paper — it's your Markdown files — but a self-updating graph tied to app-specific linking logic isn't a one-click export once you've lived in it a year.