Connect Claude Code or Cursor
Avert’s MCP server exposes the named domain tools. Each client connection is stored as a Session in the
graph and linked to the findings it touched, so a second client can recall what the first one decided.
The hosted demo (read-only, nothing to install)
Section titled “The hosted demo (read-only, nothing to install)”claude mcp add --transport http avert-demo https://avert-web.bravetree-b234f308.uaenorth.azurecontainerapps.io/mcpIt serves pre-indexed graphs with the read tools only: the synthetic avert:acme service and avert:flutter-stripe__flutter_stripe,
plus three small public Python repos with real findings (avert:whapi-cloud__python-whatsapp-chatbot, avert:kentemman__messengergpt
and avert:alonlavian__ragent). Pass one of those names as repo. Write tools (index_repo, claim_finding, open_pr and the rest) are not exposed. Each client is limited
to 60 requests a minute across /mcp and /api; beyond that the server answers 429 with a Retry-After header.
Your own repositories (local, full tools)
Section titled “Your own repositories (local, full tools)”Avert is on npm as avert-mcp. Start FalkorDB, then register the server:
docker run -d -p 127.0.0.1:6380:6379 falkordb/falkordb:v4.22.1claude mcp add avert -e FALKORDB_URL=falkor://localhost:6380 -- npx -y avert-mcp mcpAdd --read-only after mcp to hide the write tools.
Cursor
Section titled “Cursor”.cursor/mcp.json (or the global one):
{ "mcpServers": { "avert": { "command": "npx", "args": ["-y", "avert-mcp", "mcp"], "env": { "FALKORDB_URL": "falkor://localhost:6380", "AVERT_REPO": "/abs/path/to/your/repo" } } }}MCP Inspector
Section titled “MCP Inspector”npx @modelcontextprotocol/inspector npx -y avert-mcp mcpTry the handoff between two clients
Section titled “Try the handoff between two clients”scripts/act3.ts connects as two differently named clients (claude-code, then cursor). The first records a defer decision;
the second calls recall with no hints and gets it back:
pnpm tsx scripts/act3.ts examples/acmeA good first prompt
Section titled “A good first prompt”Call
recall, thenrank_findings. For the top finding, callblast_radiusand tell me which route and team it reaches.