Field notes on AI-native SaaS.
How to ship MCP servers, embed in-product AI assistants, and monetize AI usage — written by the team building the platform.
- AI monetization
Metered billing for AI agents: the worst case we priced for was below our median round
Our flat per-reply credit was floored on a 16,000-token round. Across 1,000 production turns the median round carried 19,607 prompt tokens and the heaviest 273,187. What a per-turn price actually has to bound.
Oct 10, 2026 · 7 min - MCP education
220,000 prompt tokens a turn: we stopped sending every MCP tool to the model
An 84-tool MCP deployment cost 220k prompt tokens per turn, because every bound tool schema is re-sent on every round. Per-turn tool retrieval cut schema tokens by ~75%, and broke three times before it worked.
Oct 9, 2026 · 7 min - MCP education
Capping each tool result did nothing: an agent turn needs one budget
Seven tool results that each passed an 8,000-character cap shipped 56,000 characters at the model and produced a cut-off answer. The shared per-turn budget, why water-filling beats an equal split, and the tool list you pay for on every round.
Oct 5, 2026 · 7 min - In-product AI
The assistant said it created the draft. It never called the tool.
A production fabricated write, its three root causes, and the check that catches it: why a two-word command got classified as small talk, why a bare 'yes' carries no intent, and how to detect a completed-action claim without flagging a plan.
Sep 30, 2026 · 8 min - MCP education
How does AI function calling and tool execution work?
The model never executes anything: it emits JSON, your code runs the tool, and you append the result and call the model again. The full loop in 20 lines, plus the five failure modes that only show up in production.
Sep 28, 2026 · 8 min - In-product AI
How do I prevent an AI agent from taking actions on the wrong account?
Four controls that make wrong-account actions impossible rather than unlikely: identity known by construction, an identity cache keyed on the credential subject, approvals that replay frozen arguments, and a 403 classified as a wrong-id signal.
Sep 25, 2026 · 7 min - Comparison & discovery
The best AI assistant widget to embed in a SaaS product
An honest shortlist: kapa.ai for enterprise docs Q&A, Chatbase for multi-channel website chat, Fin for ticket resolution, Command+K for an assistant that acts as the signed-in user. Plus the identity question and the pricing shapes that decide it.
Sep 24, 2026 · 7 min - MCP education
Intercom's MCP server: two regions and a registration endpoint that says no
Intercom runs separate MCP hosts for US and EU workspaces, still answers on its deprecated /sse path, uses the older OAuth discovery pattern and rejects dynamic client registration. Measured against the live endpoints.
Sep 16, 2026 · 7 min - MCP education
Slack's MCP server is live. Your users still can't connect their own workspace.
Slack's MCP server uses user tokens, supports PKCE and returns a spec-correct 401. It also has no registration endpoint, and Slack permits MCP only for Marketplace-listed and internal apps. What that means if you embed an assistant in someone else's product.
Sep 15, 2026 · 7 min - MCP education
AI agents need different docs: writing MCP tool descriptions that don't hallucinate
Agents can't ask follow-ups — your MCP tool descriptions are their entire world. Seven rules (verb accuracy, enums verbatim, ID provenance, concrete return shapes) and the AI pipeline that scales them to 84 tools.
Jul 11, 2026 · 5 min - In-product AI
Agents vs assistants vs copilots: pick the right shape
The three terms point at different autonomy levels. Picking the wrong one over-promises to users and under-builds the product. Here's the clean breakdown.
Jul 3, 2026 · 4 min