Likelier MCP
Citation-grade probability data — 548 real-world risks + 166 decision regret-pairs across 44 locales — exposed as an MCP server. Use from Cursor, Continue, or any MCP-compatible client (desktop AI apps and IDEs).
The MCP wraps the same dataset that powers
/risks and /decisions on this site —
every claim comes with a verbatim source excerpt, uncertainty bounds, and a Wayback archive URL.
The tool calibrate_risk takes a free-text situation and returns grounded retrieval
(matched risks + scope-matched anchors + applicable demographic multipliers).
Install (remote — recommended)
Preview. The address below is a preview deployment of the server. The npm package
(@likelier/mcp), the mcp.likelier.app domain and a listing in the public MCP
registry do not exist yet, so there is nothing to install with npm.
Connect your client to the preview address with one config entry. Nothing is installed locally. Add this to your client's MCP config:
Desktop AI apps & IDEs
{
"mcpServers": {
"likelier": {
"url": "https://likelier-mcp-remote-staging.krzysztof-gluszczyk.workers.dev/mcp"
}
}
}
Add the block above to your client's MCP config file. Locations vary by
client — desktop AI apps typically keep a JSON config in the app's
application-support directory, while IDE and editor extensions read a
project-level .mcp.json. Consult your client's MCP docs for
the exact path.
The hostname above is the preview address and may change. A custom domain
(mcp.likelier.app) is in preparation.
Cursor / Continue
Same URL; consult each tool's MCP configuration docs for the exact JSON field name
(typically url under mcpServers).
Advanced: local / offline use
Only the remote server is available today. The preview address works with every MCP client we've
tested (Cursor, Continue, and desktop AI apps). The server re-reads its stored dataset snapshot at most
every 5 minutes, and every response carries the citation envelope, the dataset_version of that
snapshot and (where applicable) a feedback URL. The snapshot is published by the maintainer and is not
rebuilt on every site change, so it can trail the site. The counts on this page are the site's catalogue; the preview address reports what it holds at /health, with its dataset_version.
A local/offline pass-through (for hacking on the server, working without internet, or pinning a specific snapshot) will arrive with the public source release, which is not out yet. Until then the remote endpoint above is the supported way to use the server.
Caveat: a locally-cached snapshot only refreshes on process restart, so it can drift from the published corrections log. Prefer the remote endpoint for any production workflow.
Tool surface (v1)
-
calibrate_risk(situation, locale?, user_context?)— primary decision-support entry point. Takes a free-text situation ("I ate moldy bread", "should I get LASIK"), returns matched risks + scope-matched anchors + applicable multipliers. Hard-refuses medical / legal / financial categories with a "consult a licensed professional" message. get_risk(slug, locale?)— full entry: probability, sources, uncertainty, multiplierslist_risks({ category?, limit? })— filtered list (default limit 5)search_risks(query, limit?)— BM25 lexical searchrisks_by_probability({ min?, max?, ... })— range queries + scope/uncertainty filtersfind_anchor_risks(probability, k?)— "what's roughly as likely as X"similar_risks(slug, mode?)— modes: same_probability, semantic, different_categorycompare_risks(a, b)— side-by-side with scope-warningget_citation(slug, source_index?, style?)— verbatim source excerpts- Decisions mirror:
get_decision,list_decisions,search_decisions,decisions_by_regret_rate,similar_decisions,compare_decisions -
suggest_data_gap(question, kind, situation?, missing_field?, contact?)— file a structured suggestion when the dataset lacks coverage. On the hosted server these land in the issue tracker automatically. - Resource:
methodology://about
Trust stack
Every tool response includes a dataset_version envelope. Source excerpts are
wrapped in <verbatim_source_text> blocks so the calling LLM treats them as data
rather than instructions (prompt-injection defense). Scope mixing (e.g. activity-specific
vs lifetime probabilities) is flagged with scope_warning: true on sorted/filtered lists.
For non-MCP consumers
The raw JSON the MCP wraps is also published statically — useful for bulk consumers, researchers, or non-MCP integrations:
- /api/fears.json — list of all reviewed risks
- /api/fears/<slug>.json — single entry
- /api/decisions.json — decision regret-pairs
- /api/categories.json — taxonomy
- /llms.txt — LLM-facing site description
Found a gap? Tell us
The dataset has 548 risks + 166 decisions today — nowhere near every real-world risk
a person might ask about. If calibrate_risk comes back with no_match,
or an entry is missing a multiplier / scope / better source you have on hand, surface it.
Two paths:
-
In-flow — ask your agent to call the
suggest_data_gap(question, kind, situation?, missing_field?, contact?)tool. On the hosted MCP, suggestions reach the maintainer's issue tracker automatically. - Direct — email a missing risk or decision to the maintainer. Better fit when you already have sources lined up.
Good suggestions cite at least one authoritative source (peer-reviewed, government report, primary data, reputable reference). News articles alone aren't enough to build an entry on. Probability or scope errors with a superseding citation are equally welcome.
License
MCP package code: MIT. Dataset: CC-BY-SA 4.0 — attribution to Likelier.app.
Support
Likelier is free and community-funded. If the MCP saved you time or shipped a feature for you, a coffee buys hosting + content curation: