EBP Directory

A searchable directory of evidence-based therapies and interventions for therapists and social workers, cross-referenced from major public registries. One dataset, two front doors:

• Web app (web/) — instant search + filters by problem, population, modality, and evidence tier

• MCP server (mcp-server/) — the same data served to any MCP-capable LLM (Claude Desktop, Claude Code, ...)

870 interventions · 1,011 registry ratings · 4 sources (as of the date in data/interventions.json).

Why

The registries already exist, but they're fragmented and use incompatible rating scales — the same intervention appears in four places with four different labels. This project normalizes them into one crosswalk: each intervention keeps every registry's original rating (with a provenance link), plus a normalized evidence tier so results are comparable at a glance.

Related MCP server: aria-mcp-server

Data sources

Normalized evidence tiers

Grounded in the APA Division 12 / Chambless criteria (see Royse, Program Evaluation: An Introduction to an Evidence-Based Approach, ch. 1):

An intervention's headline tier is its best rating; a ⚠ caution flag appears if any registry reports mixed/null/concerning findings (often for a specific problem or population — read the evidence rows).

Run the web app

npm run serve
# open http://localhost:8321

Or push to GitHub with Pages enabled — .github/workflows/deploy-pages.yml publishes web/ automatically.

Hook up the MCP server

cd mcp-server && npm install

Claude Desktop (claude_desktop_config.json) or any MCP client:

{
  "mcpServers": {
    "ebp-directory": {
      "command": "node",
      "args": ["C:\\Users\\above\\ebp-directory\\mcp-server\\index.js"]
    }
  }
}

Claude Code: claude mcp add ebp-directory -- node C:\Users\above\ebp-directory\mcp-server\index.js

Tools: search_interventions (query + tier/problem/population/modality/registry filters), get_intervention (full evidence with provenance links), list_filters.

Keeping it up to date

Every entry carries the registry's rating, a source URL, and the retrieval date — staleness is visible, never silent, and the registries remain the source of truth.

• Manual: python scripts/refresh.py re-scrapes everything, rebuilds the dataset, and writes a reviewable diff to data/CHANGES.md (added/removed interventions, tier changes). Review, then commit.

• Automated: .github/workflows/refresh-data.yml runs quarterly and opens a PR with the change report as its body. Nothing merges without human review — "CEBC downgraded this program" should always get eyeballs.

• A guardrail fails the refresh if the new dataset shrinks >30% (a scraper silently breaking should never wipe the data).

Pipeline

scrapers/scrape_*.py   -> data/sources/<registry>.json   (raw per-registry records)
scrapers/build_dataset.py -> data/interventions.json     (merged crosswalk)
                          -> web/data/interventions.json (copy the app loads)

Merging is by normalized name plus an alias table (ALIASES / CANONICAL_NAMES in build_dataset.py) — e.g. "TF-CBT", "Trauma-Focused CBT", and "Trauma-Focused Cognitive-Behavioral Therapy" are one entry. To fix a bad merge or add an alias, edit those tables and run npm run build-data.

Roadmap (Tier 2 sources)

• SAMHSA EBP Resource Center — blocks non-browser fetches; needs a browser-automation pass

• Title IV-E Prevention Services Clearinghouse, Blueprints, HomVEE, The Community Guide, WSIPP (effect sizes + benefit-cost ratios)

• VA/DoD and NICE guideline recommendations as evidence citations

Disclaimer

This is a reference tool that aggregates and links to public registries. It is not clinical advice. Ratings reflect the retrieval dates recorded in the dataset and may lag the registries. Registry descriptions are not reproduced wholesale; entries link to the authoritative source. Use professional judgment for individual clients.