// teleport guide · Knowledge · official

Teleport Memory (Knowledge Graph) MCP — read-only, P2P, zero cloud

The official Memory server persists an entity/relation graph in a JSON file. Teleport the read tools to let any agent ground its answers in your graph, and gate the mutating tools behind --confirm. This guide maps all 9 tools of @modelcontextprotocol/server-memory by blast radius, then gives you the exact command to share it with a remote teammate's AI agent — capability without credentials.

3SAFE to expose

3CAUTION

3KEEP LOCAL

0servers involved

host terminal — read-only share command

npx -y mcpwire share --allow "read_graph,search_nodes,open_nodes" --redact-secrets --ttl 30m -- npx -y @modelcontextprotocol/server-memory

Only the 3 read-only tools are exposed; the other 6 are invisible to the peer (filtered from tools/list and blocked if forced). Tunnel dies in 30 minutes. Add --confirm to approve every call, or --max-calls 20 for a hard budget.

peer config — claude_desktop_config.json / .cursor/mcp.json

{
  "mcpServers": {
    "teleported-memory": {
      "command": "npx",
      "args": [
        "-y",
        "mcpwire",
        "join",
        "wire://<TICKET_FROM_HOST_BANNER>"
      ]
    }
  }
}

🛡️ tool risk map — Memory (Knowledge Graph)

tool	risk	notes
`read_graph`	SAFE	Full graph dump.
`search_nodes`	SAFE	Fuzzy node search.
`open_nodes`	SAFE	Reads specific nodes.
`create_entities`	CAUTION	Writes to the graph.
`create_relations`	CAUTION	Writes relations.
`add_observations`	CAUTION	Appends observations.
`delete_entities`	HIGH RISK	Destroys graph data.
`delete_observations`	HIGH RISK	Destroys observations.
`delete_relations`	HIGH RISK	Destroys relations.

🔒 Why not just share your token?

Memory (Knowledge Graph)'s credential would grant the full toolset to a foreign LLM — including delete_entities and delete_observations. With mcpwire the credential stays in your env; the peer only ever holds an expiring wire:// capability scoped to 3 tools.

🎥 Audit everything

Add --record session.wirelog and every call, block and redaction is logged locally. Drag the file into mcpwire.co/replay for a DevTools-style waterfall of what the remote agent actually did.

🔁 Works with every MCP client

Claude Desktop, Cursor, Windsurf, Cline — anything speaking MCP over stdio. Your peer pastes the snippet above; you keep the kill-switch (Ctrl+C) in your terminal.

What happens if the peer's agent tries to call a blocked Memory (Knowledge Graph) tool?

The request never reaches your MCP server. mcpwire answers with a JSON-RPC error (-32601 [MCPWire Firewall]) and logs the attempt — the tool doesn't exist from the agent's point of view, so well-behaved agents won't even try.

Can I expose the caution-level tools too?

Yes — just add them to --allow or run with --confirm so each call waits for your y/n in the host terminal. The risk map above tells you exactly what you'd be signing up for.

Is the traffic really serverless?

Yes. Discovery goes through the public Hyperswarm DHT; the data path is a direct Noise-encrypted socket between the two machines (UDP hole punching). If both NATs are stubborn, Hyperswarm routes through community relays — which only see ciphertext.

← all teleport guides ⚡ open the policy builder

🔒 MCPWire Enterprise & Team Mesh (Early Access)

Need persistent wire://your-company/db namespaces, SSO / GitHub identity verification before the P2P handshake, and centralized audit logs for compliance (SOC2 / ISO)?