The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Fulcru listing page.
Get named by AI search engines.
When someone asks ChatGPT "what's the best CRM for a small team", the answer names two or three products. Those get the customer. Everyone else is invisible.
fulcru finds the questions where an AI assistant names a competitor instead of
you, writes the page that closes the gap, and measures what publishing it did. It
runs as a CLI, an MCP server, and an agent skill, so a human or an agent can run
the whole loop.
This repo is the open, self-hostable agent surface. The measurement engine is the hosted Fulcru service; everything here is a thin client that talks to it with your token.
We run it on ourselves and publish the result, whichever way it goes. Today it scores us 0 out of 34 prompts: we sell AI visibility and we score zero on our own product, with the full run and the source data.
Or run it in Docker, with nothing installed on the host:
No secrets are baked into the image; the token is passed at run time. Verified on
node:20-alpine, 2026-07-27.
That last line is the whole point.
| Command | What it does |
|---|---|
fulcru gaps [n] | Questions where AI names a competitor and not you, worst first |
fulcru write <promptId> | Writes the page that closes one gap; Markdown to stdout |
fulcru publish <pageId> <url> | Marks it live and snapshots the baseline to measure from |
fulcru delta | What your published pages actually did |
fulcru report [section] | overview, competitors, sources, mentions, prompts |
Add --json to any command for the raw payload.
Output is terse lines, not JSON blobs, because the caller is usually an LLM and
its context is the scarce resource. fulcru gaps costs an agent about 50 tokens;
the same data as raw API JSON costs about 800. fulcru write sends the Markdown
to stdout and everything else to stderr, so fulcru write <id> > page.md gives a
clean file.
Two ways to wire the five Fulcru tools into an agent.
Remote (hosted). Point any MCP client at the hosted endpoint with your token:
Local (stdio). For hosts that prefer a subprocess, this package ships a stdio MCP server that proxies to the hosted endpoint:
The -p is not decoration. This package ships two commands, and without it npx
runs the one named after the package, which is the CLI: the host gets a usage
screen instead of an MCP server.
Tools: fulcru_gaps, fulcru_write_page, fulcru_publish_page, fulcru_delta,
fulcru_visibility.
Both speak protocol revision 2026-07-28 and the older handshake-based
revisions, and pick per request from what the client sends. A client on the new
revision gets server/discover, per-request _meta, and a cacheable tool list;
one still sending initialize is served exactly as before. Nothing needs
configuring either way.
The skills/fulcru skill runs the whole playbook using
nothing but web access — audit a domain, find the gaps, write the page, re-measure
— with no account and no token. Install it into an agent that supports skills:
Hermes scans third-party skills before installing and prints the verdict, so you can see what you are agreeing to run. If you would rather track this repo as a source and pick up updates, tap it instead:
The CLI and MCP above are what you use when you want it measured continuously across every engine, with the before/after delta tracked for you.
| Env var | Required | Default | Purpose |
|---|---|---|---|
FULCRU_TOKEN | for the CLI + local MCP | — | Your account token (Settings -> Integrations) |
FULCRU_ENDPOINT | no | https://little-orca-977.convex.site/mcp | Point at a self-hosted or custom-domain endpoint |
No secrets are baked into this repo. The token is read from the environment only.
The skill needs nothing. The CLI and MCP need a Fulcru account for the continuous,
measured loop; point FULCRU_ENDPOINT at your own instance if you run one.
A Docker image is provided:
Issues and PRs welcome. The tool logic lives in the hosted service; this repo is the CLI, the stdio MCP proxy, and the skill. If AI names a competitor instead of you, open an issue with the question — that is exactly the gap this is built to close.
If this is useful, a ⭐ genuinely helps.
AGPL-3.0-or-later. See LICENSE.