Rowboat is an open-source AI coworker backed by Y Combinator. It connects to your email, calendar, and meeting notes, builds a knowledge graph of everything, and uses AI agents to help you act on it — draft documents, prepare for meetings, follow up on decisions, monitor topics. Think of it as an AI teammate that remembers everything and can actually do things.
Under the hood, Rowboat runs multiple specialized agents. One searches your knowledge graph. Another drafts content. Another plans tasks. Another monitors topics over time. The problem: all these agents typically use the same model. Your research agent, your content drafter, your quick-reply agent — all running on GPT-4o, paying GPT-4o prices, even when simpler models would work fine.
ANT changes this. Connect Rowboat to ANT as an OpenAI-compatible provider, and each agent can use a different virtual model. The research agent gets deep reasoning. The quick-reply agent gets sub-300ms responses. The background monitor runs on free models. Same capability, 60-70% lower cost.
What Rowboat Does
Rowboat turns your work into a living knowledge graph:
- •Connects to email, calendar, and meeting notes
- •Builds a persistent knowledge graph of people, companies, projects, and decisions
- •Live Notes — automatically updated notes that track topics across all your sources
- •Natural language commands — "Build me a deck about our Q3 roadmap" or "Summarize what happened in the engineering standup"
- •MCP tool support — connects to external tools via Model Context Protocol
- •Local-first — all data stored as Markdown in an Obsidian-compatible vault
- •Voice input/output via Deepgram and ElevenLabs
Why Each Agent Needs a Different Model
A typical Rowboat session involves multiple agents working in sequence:
- •Knowledge retrieval agent — searches your graph for relevant context. Needs speed, not deep reasoning. → ant:fast (Gemini Flash, $0.30/1M)
- •Analysis agent — synthesizes information from multiple sources, finds patterns. Needs reasoning. → ant:reasoning (o3, DeepSeek R1, $2-10/1M)
- •Content drafting agent — writes emails, docs, presentations from the analysis. Needs quality. → ant:coder or ant:best (Claude Opus, GPT-5, $3-15/1M)
- •Quick reply agent — handles simple questions, confirmations, status checks. Needs speed. → ant:fast ($0.15/1M)
- •Background monitoring agent — watches for new mentions, updates live notes. Runs continuously. → ant:free (Llama, Gemma, $0/1M)
- •Meeting prep agent — compiles relevant context before meetings. Needs thoroughness. → ant:reasoning ($2-10/1M)
Without ANT, all six agents use GPT-4o at $2.50-10/1M tokens. With ANT, only the analysis and content agents hit premium models. The monitoring agent runs for free. The quick-reply agent uses a $0.15 model. Total savings: 60-70%.
Setup: Connect Rowboat to ANT
Rowboat supports OpenAI-compatible providers. Add ANT to your configuration:
// ~/.rowboat/config/models.json
{
"providers": [
{
"name": "ant",
"flavor": "openai-compatible",
"apiKey": "your-ant-api-key",
"baseURL": "https://antbase.ai/v1",
"models": [
{ "id": "ant:auto", "name": "ANT Auto (best for each task)" },
{ "id": "ant:coder", "name": "ANT Coder (content generation)" },
{ "id": "ant:reasoning", "name": "ANT Reasoning (analysis)" },
{ "id": "ant:fast", "name": "ANT Fast (quick replies)" },
{ "id": "ant:free", "name": "ANT Free (background tasks)" },
{ "id": "ant:best", "name": "ANT Best (critical content)" }
]
}
]
}What to Build with Rowboat + ANT
Example 1: AI Sales Assistant
Before every sales call, Rowboat's knowledge graph has context from previous emails, meeting notes, and CRM data. The meeting prep agent (ant:reasoning) synthesizes a briefing. During the call, the quick-reply agent (ant:fast) suggests responses in real-time. After the call, the content agent (ant:coder) drafts a follow-up email. The monitoring agent (ant:free) watches for the prospect's reply.
Example 2: Engineering Team Coordinator
Rowboat connects to GitHub, Slack, and your calendar. The analysis agent (ant:reasoning) identifies blocked PRs and overdue tasks. The content agent (ant:coder) drafts standup summaries. The monitoring agent (ant:free) watches for CI failures and pings the right person. The quick-reply agent (ant:fast) handles "what's the status of X?" questions from stakeholders.
Example 3: Research & Writing Pipeline
You say: "Write a report on our competitive landscape." Rowboat's retrieval agent (ant:fast) pulls everything from your knowledge graph. The analysis agent (ant:reasoning) identifies trends and gaps. The content agent (ant:best) writes a polished report with citations. The monitoring agent (ant:free) sets up a live note to track future competitive developments.
Cost Comparison
A typical Rowboat workflow processing a complex request (like preparing a meeting brief):
- •All agents on GPT-4o: 8 agent calls × ~$0.05 avg = $0.40/task
- •With ANT routing: retrieval on ant:fast ($0.002) + analysis on ant:reasoning ($0.08) + drafting on ant:coder ($0.05) + monitoring on ant:free ($0.00) = $0.13/task
- •Savings: 68% per task
- •At 100 tasks/day: $40/day → $13/day = $810/month saved
Automatic Failover
Rowboat's agents run continuously — monitoring live notes, watching for updates, preparing context. If a single model provider goes down, the whole system stalls. With ANT, each agent has automatic failover across 30+ providers. Anthropic down at 3 AM? The content agent switches to GPT-5. Gemini rate-limited? The retrieval agent moves to Mistral. Your AI coworker never sleeps and never stops working.
The Knowledge Graph + Routing Intelligence
This is where the combination gets interesting. Rowboat builds a knowledge graph that grows over time — every email, every meeting, every document adds context. ANT routes queries about that knowledge to the optimal model based on complexity. A simple lookup ("When is my next meeting with Sarah?") goes to a fast, cheap model. A complex analysis ("What decisions from the last 3 board meetings are still unresolved?") goes to a reasoning model. The intelligence is in both layers — Rowboat knows what to ask, ANT knows who to ask.
Get Started
1. Clone Rowboat from GitHub (github.com/rowboatlabs/rowboat) 2. Add ANT as a provider in models.json 3. Get your API key at antbase.ai ($5 free credits) 4. Start building — each agent automatically gets the right model
