Why We Don't Put AI in Your Feedback Board
Every feedback tool is adding an AI tier. We built an MCP server instead — so your own agent reads your board, and the cheap parts stay free for everyone.
Every feedback tool in this category now has an AI tier. Canny has Autopilot. Featurebase has Growth AI. Sleekplan has Sleek Intelligence. featureOS bundles AI analytics into a $60/month suite. The pattern is consistent enough to be a category norm: add a model, add a tier, add a price.
FeatureWish doesn’t have one, and it isn’t an oversight. We looked at building it — there’s a costed plan in our repo for exactly this, local embeddings and all — and decided against it. Here’s the reasoning, because “we don’t have the AI feature” deserves a real answer rather than a shrug.
Two different things get called “AI in your feedback tool”
The conversations get muddled because one label covers two jobs that have almost nothing in common.
The cheap job is catching a request that’s already been posted. Someone types “dark mode please” when eleven people have already asked for dark mode. You want that caught at the keyboard, before it becomes a twelfth post splitting the vote count twelve ways.
The expensive job is judgment. What should we build next? What are these two hundred requests really about? Is this worth a week? That’s not pattern-matching, it’s reasoning about your product, your users, and your constraints.
Most tools price these together and put both behind the same upgrade. They’re not the same, and they shouldn’t be.
The cheap job shouldn’t cost extra
Catching restated requests barely needs a model. FeatureWish does it with trigram matching in Postgres — as someone types a title, similar existing requests surface underneath, and they upvote instead of reposting. It’s typo-tolerant, it runs in the database, and it costs us close to nothing per query.
So it’s free. On every plan, with no AI add-on, because charging for it would mean charging you to keep your own board tidy.
Here’s the honest limitation: trigram matching compares character patterns, so it catches “dark mode” against “darkmode” or “dark theme please” — restatements and typos, which is most duplicates in practice. It will not catch “export to spreadsheet” against “CSV download.” Those share almost no characters. A semantic model would catch it; ours doesn’t.
We think that trade is worth it at the price of $0, and we’d rather tell you where the line is than imply there isn’t one. Manual merge is there for the rest, and it carries votes and comments across.
The expensive job shouldn’t come from us
This is the part where we differ from the category, and it’s the part that actually matters.
An AI assistant living in a feedback dashboard knows exactly one thing: your feedback. That’s it. It doesn’t know your codebase, so it can’t tell you the third-most-requested feature is four days of work and the seventh is an afternoon. It doesn’t know you’re two months from running out of money. It doesn’t know you tried something similar last year and it flopped. It’s reasoning about your product through a straw.
Meanwhile you already have an assistant that knows all of that. It’s the one you code with.
So we didn’t build an assistant. We built the connection — a Model Context Protocol server, so Claude, Codex, Cursor, or any MCP client can read your board directly. Seven scoped tools: list boards and posts, read a request with its full comment thread, pull the roadmap and changelog, file a post, update a status.
The difference in practice isn’t subtle:
“Look at the top ten open requests on my board. For each one, skim the codebase and estimate rough effort. Which three have the best votes-to-effort ratio?”
No dashboard chatbot can answer that, because half the inputs live in your repository. Your agent can — and it shows its reasoning, so you can push back on it.
The same goes for the rest of what usually ships in an AI tier. Summaries: “read the CSV export thread and tell me what people actually need.” Clustering: “group these two hundred requests into themes.” Changelog drafting: “we merged dark mode — mark it completed and draft an entry from the PR.” That last one flips the request to Completed and emails every voter automatically, from your editor, without a browser tab.
You don’t need our AI. You need your AI to see your feedback.
What this means for your bill
The practical consequence is that there’s no credit meter here. No AI tokens to top up, no usage that turns into an invoice surprise, no tier you get pushed onto because your board got popular.
The dedup layer is free. The MCP server is on Pro at $20/month flat — the same flat price as everything else, never metered, alongside your workspaces, boards, GitHub and Linear sync, custom CSS, and white-labeling. You bring your own agent, and you’re already paying for it.
We’ll be honest about the counterargument: if you want AI analysis inside the tool, on every plan, without wiring anything up, tools like featureOS and Featurebase do that today and we don’t. If you have no AI assistant in your workflow at all, our answer asks something of you that theirs doesn’t.
But if you’re already building with an agent — and most teams shipping software in 2026 are — then a chatbot bolted onto a feedback dashboard is a strictly worse version of a thing you already own.
Create a free board, or read how founders actually use an agent-readable board.