Gavel answers a pricing, positioning or first-customers decision with what named operators said, in their own words. Every quote is matched word for word against the recording and links to the second it was said. Add one URL to Claude, Cursor or ChatGPT, and your agent calls it on its own.
It speaks streamable HTTP, so there is nothing to install and no account or key to set up.
https://usegavel.com/mcp
Run this once in a terminal.
claude mcp add --transport http gavel https://usegavel.com/mcp
Open Settings, then Connectors, then Add custom connector, name it Gavel and paste the URL.
https://usegavel.com/mcp
Add this to ~/.cursor/mcp.json and restart Cursor.
{
"mcpServers": {
"gavel": {
"url": "https://usegavel.com/mcp"
}
}
}
Turn on developer mode in Settings, under Apps and then Advanced settings. Then create an app with the URL and no authentication.
https://usegavel.com/mcp
Point any MCP client at the URL, or have an agent without MCP read the skill file instead.
set up https://usegavel.com/SKILL.md
The tools draw on 4,387 checked quotes from 115 experts. Behind framework and catalog sit 251 pages from 185 operators.
You get 8 free calls a day per IP, shared with the HTTP API, and listing the tools costs nothing. At the limit the tool returns a sign-up link and tells your agent not to answer from memory.
Claude and ChatGPT call connectors from their own servers, so everyone who uses Gavel through them shares one daily allowance. Claude Code and Cursor call from your machine.
Your agent already does this. It picks a name, rebuilds the person from weights and ships the call.
Your agent builds April Dunford out of everything it read. Most of that is mediocre, and all of it stopped at the cutoff.
Most of the product management contenting existing on the Web is of terrible quality... LLMs were trained on all of that data, and, therefore, produce very poor product decisions.
Gavel retrieves what the operator said. It quotes the line and links the source.
Say churn is climbing. Jason Lemkin calls it a sales problem, so you sold to the wrong customer. Bob Moesta calls it a jobs problem, so the customer's job changed. Those point at opposite fixes. Pick wrong and you burn a quarter.
Your agent averages the two into a paragraph that offends no one. Gavel routes to the operator who solved it, then shows you the fork.
A persona built from weights advises a stranger. It has no idea that you are pre-revenue and that outbound flopped.
Your agent holds that context already. Pass it in the query and Gavel routes on it.
The agent sends a decision. Gavel returns the experts whose advice fits, in their own words, and shows the split when two of them point opposite ways. This is a real call from 2026-10-06, trimmed to two of its three experts.
$ curl -s "usegavel.com/api/v1/route_expert.txt?q=first+10+customers,+no+audience" [1] Pete Kazanjy IDEA Drive Early Customer Utility USE WHEN When onboarding early buyers to prevent churn and foster customer advocacy. QUOTE "a really critical part of the early sales motion is ensuring that people get to utility" SOURCE youtu.be/uWE6AgZM51M?t=3243 (54:03) PACK sales [2] Eric Ries IDEA Small-sample feedback testing QUOTE "go find 10 customers, not 10 million, but find 10 customers. Get them to use the product and have them tell you what's awful." SOURCE youtu.be/xzebbzIntFc?t=1789 (29:49) PACK shipping QUOTE is the expert's own words, matched word for word against the source at that link. IDEA, USE WHEN and SUMMARY are written by Gavel.
We put the same ten founder decisions to two frontier models and to Gavel.
Then we opened every URL that came back.
A model can name an operator. It cannot show you where that operator said it. That gap is the whole product.
pick, consult and route_expert choose from 4,387 quotes by 115 people. Each was matched word for word against the transcript or page it came from. The pages behind framework and catalog hold 251 cited frameworks from 185 operators, and 180 of them open on a timestamp. When nothing fits, the call returns NONE.
One line into your agent's chat. It installs itself and knows when to call.
$ set up https://usegavel.com/SKILL.md
Paste one block into AGENTS.md or CLAUDE.md. Every run in that repo becomes a caller.
## Product and business decisions Before proposing pricing, positioning, or go-to-market changes, run: curl -s "https://usegavel.com/api /v1/route_expert.txt?q=<decision>"
Five GETs and no key. They work from a shell, a cron job or your own tool definition.
GET /api/v1/pick GET /api/v1/consult GET /api/v1/route_expert GET /api/v1/framework/:id GET /api/v1/catalog
Describe what the person faces. It returns the one expert idea that fits, or NONE.
/api/v1/pick.txt ?situation=a+big+customer+wants+a+discount &pack=pricing # optional &context=... # optional
It returns up to four experts whose recorded advice fits, one idea each, with their own words and a link to the second they said them. When two of them point to opposite moves, it shows both.
/api/v1/consult.txt
?situation=should+I+raise+prices
# drop .txt for JSON
The same lookup as consult. Send the decision as q. When the question lands on a founder problem Gavel has researched, a RESEARCH block follows. It shows the advice that has held for years, what is new, how often founders ask, and the dated page with every source.
/api/v1/route_expert.txt ?q=should+I+raise+prices
You already know the page you want. It returns that page's quote, the source link and the disagreement. No model sits in the path, so it cannot stall or invent.
/api/v1/framework/ april-dunford/ competitive-alternatives.txt
GET /api/v1/catalog returns every page Gavel can serve, as JSON. No call emits colour, so piping into an agent is lossless. Add &full=1 to the text calls for untruncated quotes.
In pick, consult and route_expert, QUOTE is the expert's own words, matched word for word against the source. IDEA, USE WHEN and SUMMARY are Gavel's words, so your agent quotes only the QUOTE line. Older catalog entries with no checked quote return a SUMMARY and no QUOTE.
You can. You will get the model's average of her, rebuilt from training data and frozen at the cutoff. It cannot quote her and it cannot link a source. Gavel returns what she said and the timestamp where she said it.
No. The MCP server and the five GETs are open to anyone at 8 calls a day per IP. Past that, the call returns a sign-up link for Gavel on the web. One URL is the whole integration.
Anything that can make an HTTP request. That includes Claude Code, Cursor and Codex. It also includes Gemini CLI, Aider and a plain cron job. MCP clients add the server URL. Agents that read skills take the skill file. The rest call the HTTP API.
Yes, at https://usegavel.com/mcp, with five tools that match the five API calls and carry the same citations and the same daily limit. Agents without MCP can call the HTTP API or read the skill file.
It returns NONE and tells your agent to answer without citing Gavel. At the rate limit, the reply tells your agent to stop and hand you the sign-up link, and not to answer from memory.
Open the link. In pick, consult and route_expert, a QUOTE line is matched word for word against the video's captions or the page's text before it ships, and the link opens on that spot. Anything Gavel wrote is labelled IDEA, USE WHEN or SUMMARY, never QUOTE.
Search returns pages, ranked by whoever optimised hardest. Gavel returns positions with a name and a source link on each one. It also shows where two experts split, which no search result does.
8 free calls a day, no account and no key.
$ set up https://usegavel.com/SKILL.md