Nobody wanted our AI agent. They wanted its tools.
Last month we shipped an SEO agent inside Tracetify. It plans research across your Search Console, site audits and competitor reports, then answers with citations. We were proud of it.
Then we watched how developers actually work. They sit in Claude Code or Cursor all day, and those agents are often better than ours. Asking a developer to leave their editor to talk to our agent was asking them to downgrade.
So phase 2 went the other way: we exposed the data, not the agent. The MCP server grew from 5 tools to 14 this week:
- Growth traces: how any product actually got its first users, rebuilt from 12 public sources in about 60 seconds, every claim linked to a dated citation. We ran it on a competitor yesterday and learned their "established platform" published its launch PR 14 months ago.
- Your own Search Console: the queries sitting at position 5 to 20. The part I like is what happens next. The agent sees "this page ranks #12" and edits that page in the same session, because unlike our dashboard, it has your code.
- Site audits after deploy, a hand-checked list of 58 directories that give real dofollow links, and AI visibility lookups.
Two decisions that felt right:
Tools that cost credits quote the price and wait for a yes. A site audit costs 3 credits, and the agent will tell you that before running it. Nothing spends while you sleep.
Zero-install option. npx -y tracetify-mcp works, but you can also point any MCP client at https://tracetify.com/api/mcp and skip Node entirely. Reading the public report library is free either way.
The setup page has a prompt you paste into Claude Code that installs everything and runs a demo query on its own. A competitor taught us that trick. It works.
Docs: https://tracetify.com/mcp
Honest question for people building MCP servers: our free tier is "read everything already traced." Is that the right hook, or would you open one paid tool instead?
The “quote the action and wait for approval” pattern is interesting because it solves one half of the trust problem: was the agent authorised to act?
The other half starts immediately afterwards: did it make the right change?
Once an agent can inspect a live site and modify it in the same session, have you thought about independently sampling completed actions — comparing what the user authorised, what the agent actually changed, and whether the resulting state still satisfied the intended objective?
That authorised → executed → independently verified chain is something we’ve been exploring with OpsWatch.
Tracetify’s tool-level evidence could make that kind of assurance unusually clean.
That's exactly how it felt from the inside. The capability was never the problem — the dashboard agent answers the same questions fine. It just lives in the wrong place for someone who's mid-commit and doesn't want to open another tab. Once we framed it as "where does the user already sit," the roadmap kind of wrote itself.
The shift from selling the agent to exposing the tools is a meaningful product signal. The fact that developers preferred bringing the capability into their existing workflow says a lot about where the friction actually was.