is this true?

A small template showing how to use the X (Twitter) MCP from an eve agent.

Paste a tweet link or type a claim, and the agent fact-checks it against what's actually posted on X, then returns a structured verdict card. It's the "@grok is this true?" reply, as a web app.

How it works

Browser (Next.js)  ──►  eve agent  ──►  X MCP (https://api.x.com/mcp)
   single input          model loop        search posts, look up users,
   verdict card          + outputSchema     fetch tweets, news, trends
  1. The browser sends the claim with a per-turn output schema (lib/verdict.ts), so the model must return a structured verdict.
  2. The agent uses the X MCP connection (agent/connections/x.ts) to search the X archive, fetch a linked post, look up users, and read news/trends.
  3. The UI (app/_components/is-this-true.tsx) renders the verdict, confidence, and the cited X posts, with the original tweet embedded when you paste a link.

The X MCP connection

X hosts a Streamable HTTP MCP server at https://api.x.com/mcp. A deployed web app reaches it with an app-only Bearer token sent as Authorization: Bearer <token> — read-only access (search posts, look up users/tweets, news, trends), which is exactly what fact-checking needs. (X's other route, the local xurl stdio bridge, is for desktop clients like Cursor and can't run in a deployed app.)

eve just needs that token. You can give it to eve two ways, and the connection already supports both via the X_CONNECTOR env var:

How the token is providedSetup
A. Static credential (default)X_BEARER_TOKEN env varbelow
B. Vercel ConnectStored in Connect, fetched at runtime (no secret in your env)docs/vercel-connect.md

Either way the token never reaches the model — eve resolves it per step and the model only sees the discovered tools as x__<tool>.

Prerequisites

  • Node 24+
  • An X developer app with an app-only Bearer token (X Developer Portal → your app → Keys and tokens → Bearer Token).
  • A model credential for the eve agent (routes through the Vercel AI Gateway): either AI_GATEWAY_API_KEY, or a VERCEL_OIDC_TOKEN from vercel link && vercel env pull.

Setup

npm install
cp .env.example .env.local

Then pick one of the two options below and set the matching env var(s).

Option A: Static credential (default)

Put your X Bearer token in .env.local:

# .env.local
X_BEARER_TOKEN="AAAAAAAAAAAA...your X app-only Bearer token..."

Paste the token exactly as the portal shows it. Don't URL-decode it — X accepts the verbatim value, and decoding %2B/%3D yields a 401.

That's it. The connection (agent/connections/x.ts) reads X_BEARER_TOKEN and sends it as the Bearer token.

Option B: Vercel Connect

Store the token in Vercel Connect instead, so no secret lives in your env/repo. For this read-only app the right Connect primitive is an API-key connector (app-scoped, non-interactive). Full walkthrough:

→ docs/vercel-connect.md

Short version: create an API-key connector holding your Bearer token, then set X_CONNECTOR="api.x.com/x-key" instead of X_BEARER_TOKEN. The connection switches to Connect automatically.

Run it

npm run dev

Open the app, paste a tweet or type a claim, and watch it check X.

Deploy

This deploys to Vercel as one project (Next.js app + eve runtime):

# Option A (static credential):
vercel env add X_BEARER_TOKEN     # paste your token, all environments
vercel deploy --prod

# Option B (Vercel Connect): see docs/vercel-connect.md, then
vercel env add X_CONNECTOR        # value: api.x.com/x-key
vercel deploy --prod

Public demo: agent/channels/eve.ts uses none(), so the deployed endpoint is open — anyone can use it and every request spends your X rate limits and AI tokens. For more than a demo, replace none() with real auth (Auth.js, Clerk, …); see node_modules/eve/docs/guides/auth-and-route-protection.md. Also note Vercel Deployment Protection (SSO) will gate the whole site if enabled — turn it off for a truly public demo.

Make it yours

  • Narrow the X tools the model can call: run eve info to see the exact tool names, then set tools.allow in agent/connections/x.ts.
  • Change the verdict shape: edit lib/verdict.ts (the schema is shared by the agent and the UI).
  • Tune the behavior: edit agent/instructions.md.

Caveats

  • Verdicts are based on evidence found on X, not absolute ground truth — the X API tells you what's on X, not what's objectively correct.
  • The app-only Bearer token is read-only and has no user context (it can't post).
  • This is an automated AI fact-check and can be wrong.