OpenEvidence MCP

Query OpenEvidence from Claude Code, Codex CLI, Antigravity CLI, Claude Desktop, Cursor, Cline, Continue, and any MCP client — authenticated through your own logged-in browser tab. No API key.

What it does

OpenEvidence protects its API with bot detection that blocks plain server requests. This project removes that problem: when your AI tool asks OpenEvidence a question, the request is run inside your own logged-in OpenEvidence browser tab — so it carries your genuine browser session and is never challenged.

A small Chromium browser extension lends its session to a localhost relay; the MCP server speaks to that relay. The extension is a generic authenticated fetch proxy — all the OpenEvidence logic stays in the local server, and your browser login is the only credential. No API key, no cookie file, no Playwright, no headless browser.

It is designed for local personal workflows where you already have lawful access to OpenEvidence. It does not bypass authentication, remove access controls, redistribute OpenEvidence content, or include any OpenEvidence data in this repository.

How it works

  Claude / Codex / any MCP client
            │  asks a question (oe_ask)
            ▼
     openevidence-mcp  ──▶  relay daemon  ──▶  browser extension  ──▶  your logged-in
       (stdio server)      127.0.0.1:8787       (runs the fetch)        OpenEvidence tab
            ▲              shared · auto-spawned                              │
            └──────────────────────── answer ◀───────────────────────────────┘

Nothing navigates or pops up — the tab stays where it is. The extension only ever talks to openevidence.com and your local relay (127.0.0.1). The relay runs as a shared daemon that owns port 8787 and outlives every session, so any number of Claude/Codex sessions funnel through the one logged-in tab.

Quick start

git clone https://github.com/htlin222/openevidence-mcp.git
cd openevidence-mcp
make all      # installs deps · builds the MCP server + relay extension ·
              # registers the server into Claude and Codex (whichever CLI you have)

Then the one manual step make all prints (a browser action that can't be scripted):

1. Load the extension — open chrome://extensions (Chrome / Edge / Brave / Arc / Vivaldi / Opera) → turn on Developer mode → Load unpacked → select extension/dist.
2. Stay logged in to openevidence.com in that browser, and keep a tab open. That login is your authentication.
3. Run your AI tool. The server auto-starts the relay and connects to the extension.

Verify and go:

curl -s http://127.0.0.1:8787/health     # expect {"ok":true,"connected":true,"version":1,...}

Then just ask, in any MCP client: “Use OpenEvidence to answer …”. Re-run make all anytime to rebuild + re-register; make help lists every target; make kill-all stops all servers + the relay daemon.

Clicking the extension's toolbar icon opens a built-in how-it-works page with a live connection check.

cookies.json and a HAR are optional — needed only for the legacy OE_MCP_RELAY_TRANSPORT=off cookie read path, the Python collections tooling, and npm run doctor / login / smoke. See Optional cookie path.

One tab, many sessions

The relay is a standalone daemon (auto-spawned; run/inspect with npm run relay) that owns port 8787 and is shared by every MCP server. You can run OpenEvidence from any number of Claude/Codex sessions at once — they all flow through the single logged-in tab. The daemon:

• outlives session restarts and is respawned automatically if it dies (a crashed daemon is replaced and the in-flight request retried);
• is idempotent — a second daemon that loses the port race exits cleanly;
• reports version + pid on /health, so an upgrade replaces a stale daemon from an older build.

Asking questions

After registration, ask your MCP client in plain English and mention OpenEvidence — the agent calls oe_ask automatically.

Use OpenEvidence to answer: DLBCL frontline treatment landscape NCCN v3.2026. Include citations and BibTeX.
Use OpenEvidence to compare Pola-R-CHP vs R-CHOP in untreated DLBCL. Include trial citations and BibTeX.
Use OpenEvidence to review current evidence for SGLT2 inhibitors in HFpEF. Include citations and BibTeX.
Use OpenEvidence to find guideline-supported anticoagulation options for cancer-associated thrombosis.

Fire-and-forget by default

oe_ask returns {article_id, status:"pending"} the moment the question is submitted, freeing the tab for other sessions instead of holding the call for the whole generation:

// oe_ask  →  returns instantly
{ "article_id": "…", "status": "pending" }

// oe_article_get(article_id)  →  the finished answer, when ready
{ "article_id": "…", "status": "success", "extracted_answer_raw": "…", "figures": […], "artifacts": { … } }

• Fetch the answer later with oe_article_get(article_id) — pass wait_for_completion: true there to block until it's ready.
• For one-shot blocking (submit and wait in a single call), pass wait_for_completion: true to oe_ask itself.

A completed article returns the OpenEvidence payload and status, the article_id, the answer markdown as extracted_answer_raw, any figures, inline BibTeX as artifacts.bibtex, and saved citation files. Pass include_bibtex: false to keep the response small while still writing citations.bib to disk.

Tools

Privacy & security

• The relay listens on localhost only (127.0.0.1:8787) — nothing is exposed to the network.
• The extension and server store no credentials. Your OpenEvidence session lives in your browser, as always.
• The extension only acts on requests from your own local relay, and only against openevidence.com.
• This repository contains connector code only — no OpenEvidence content, datasets, cookies, or account material.

Troubleshooting

• Extension badge not green / connected:false? Make sure you're logged in to openevidence.com in that browser with a tab open, then reload the extension (chrome://extensions → ↻).
• Tools fail with “relay not connected”? Start your AI tool / MCP server so the daemon comes up, then curl -s http://127.0.0.1:8787/health. If a previous build is stuck, make kill-all and reconnect the MCP server.
• Run the relay in one browser at a time if you've loaded the extension in several — requests go to whichever polls first.
• DataDome 403 on the legacy cookie path? See Doctor below — relevant only when OE_MCP_RELAY_TRANSPORT=off.

Register with MCP clients

make all already registers Claude Code and Codex CLI. To do it à la carte (or for other clients), register the local stdio server node /ABSOLUTE/PATH/openevidence-mcp/dist/server.js. No env is required — the browser extension is the login.

make install-claude-global    # claude mcp add-json --scope user openevidence …
make install-codex-global     # codex mcp add openevidence -- node dist/server.js
make install-agy-global       # Antigravity CLI (agy-cli)
make install-all              # all three

For Claude Desktop, Cursor, Cline, Continue, use this mcpServers shape (see examples/ for a full-knobs reference):

{
  "mcpServers": {
    "openevidence": {
      "command": "node",
      "args": ["/ABSOLUTE/PATH/openevidence-mcp/dist/server.js"]
    }
  }
}

Collections sync & auto-sort

scripts/collection_sort.py mirrors your chat history and collection memberships into a local SQLite (~/.openevidence-mcp/db/oe.sqlite; override with OE_MCP_DB_PATH). The routine routines/collection-sort.md walks an MCP client through syncing, surfacing unsorted chats, and applying multi-membership hashtag tags. Convention: collections whose name starts with # are agent-managed; collections without a leading hash are human-curated and never touched.

The same pipeline is exposed as the oe_collections_* MCP tools (the TS server shells out to scripts/collection_sort.py via python3, overridable with OE_MCP_PYTHON).

python scripts/collection_sort.py init
python scripts/collection_sort.py sync-history --full   # first time
python scripts/collection_sort.py sync-collections
python scripts/collection_sort.py list-unsorted --json  # routine reads this
python scripts/collection_sort.py summary

Account-scoped (schema v2). Chats, collections, and memberships carry an account column (composite primary keys), so one DB can mirror multiple OpenEvidence logins without mixing them. The account is resolved from /api/auth/me; override with --account EMAIL. A v1 DB auto-migrates on first open — pre-v2 rows are tagged --legacy-account (default legacy; or OE_MCP_LEGACY_ACCOUNT).

Schedule the sync (macOS)

The classification step needs the agent in the loop, but the sync side is pure I/O — install a daily launchd job that keeps the local mirror fresh:

bash scripts/install_launchd.sh                  # daily 02:00 (OE_MCP_SYNC_HOUR / OE_MCP_SYNC_MINUTE)
launchctl start com.htlin.openevidence-mcp.sync  # fire once to verify
tail -30 ~/.openevidence-mcp/logs/sync.log
bash scripts/install_launchd.sh --uninstall      # remove

The wrapper (scripts/collection_sync_cron.sh) appends one block per run to ~/.openevidence-mcp/logs/sync.log. It takes an optional mode flag:

scripts/classify.py runs offline (no API): a per-tag log-odds-ratio signature (Monroe et al. 2008) built from your memberships every run, OR'd with curated keyword rules. Validate with python scripts/classify.py validate. Tune headless use via OE_MCP_AUTO_THRESHOLD (default 12) and OE_MCP_AUTO_TOP_K (default 3); switch the job to autonomous mode with OE_MCP_SYNC_MODE=--auto bash scripts/install_launchd.sh.

Citation artifacts

Completed oe_ask / oe_article_get calls save artifacts under ${OE_MCP_ARTIFACT_DIR}/<article_id>/ (default OS temp dir + openevidence-mcp; on macOS, /tmp may resolve under /var/folders/.../T/):

Crossref validation: DOI citations are validated directly; non-DOI citations use a bibliographic query and are marked candidate / not_found / error. Low-similarity matches never overwrite BibTeX metadata, and sources like NCCN guidelines may stay as local OpenEvidence metadata when Crossref has no authoritative match.

Optional cookie path

The extension relay is the default and recommended path. For headless reads without the extension, set OE_MCP_RELAY_TRANSPORT=off to route reads (oe_history_list, oe_article_get, oe_collections_*) over a browser-exported cookies.json — asks still need the extension. This path also backs the Python collections tooling and the npm run doctor / login / smoke CLIs.

cp /path/to/browser-cookies.json ./cookies.json
make build HAR=/path/to/www.openevidence.com.har   # extracts the browser fingerprint, then compiles
npm run login && npm run smoke

make build extracts openevidence-fingerprint.json from the HAR when present. The fingerprint is profile-faithful — the client sends exactly the captured browser's header set, so a HAR from any browser (incl. Safari) stays coherent with the cookie that browser minted.

Doctor (legacy cookie path)

When the cookie path fails with a DataDome 403 — most often after moving to a different computer — run the doctor:

npm run doctor              # static checks + a live read probe
npm run doctor -- --offline # static checks only (no network)
npm run doctor -- --json    # machine-readable output

It flags datadome-missing / -expired / -session (cookie absent / past expiry / session-scoped), fingerprint-platform-mismatch (cookie + fingerprint minted on a different OS — re-mint here), fingerprint-default (no fingerprint; built-in signature in use), and datadome-live (a live request was actually challenged). Non-zero exit on failure for CI/pre-flight.

Make targets

Run make help for the grouped, always-current list.

Environment variables

Project files

• extension/README.md — the browser-extension relay (and its built-in README.html how-it-works page)
• README.AI.md — agent install playbook
• src/server.ts — MCP tools
• src/relay-server.ts — the localhost relay (extension-facing + /relay bridge)
• src/relay-client.ts / src/relay-daemon.ts — shared-daemon client + standalone daemon
• src/citations.ts — citation extraction, BibTeX, Crossref validation
• src/doctor.ts — stale DataDome cookie diagnostics (legacy path)
• docs/plans/2026-06-04-shared-relay-daemon-design.md — relay daemon design doc
• examples/ — MCP client config samples

Copyright, trademark, and medical disclaimer

This project is unofficial and independent. It is not affiliated with, endorsed by, sponsored by, or approved by OpenEvidence or its owners. "OpenEvidence" and related names, logos, product names, and content remain the property of their respective owners.

This repository contains connector code only. It does not include OpenEvidence copyrighted content, proprietary datasets, model outputs, article payloads, session cookies, or account material. Your local use of this MCP server may create files such as answer.md, article.json, and citations.bib; those artifacts can contain content retrieved from or derived from your OpenEvidence account session. Treat those files as private unless you have the right to share them.

You are responsible for complying with OpenEvidence terms, institutional policies, copyright law, and any clinical data governance rules that apply to your use. Do not publish cookies, account tokens, saved article payloads, generated answers, screenshots, guideline text, or other protected/copyrighted content unless you have permission or another valid legal basis.

This software is not medical advice and is not a medical device. It is an integration tool for an MCP client. Clinicians and qualified users remain responsible for verifying outputs against authoritative sources and applying independent clinical judgment.

License and attribution

Apache-2.0. Keep LICENSE and NOTICE when redistributing.

Based on OpenEvidence MCP by Bakhtier Sizhaev: https://github.com/bakhtiersizhaev/openevidence-mcp