A gatekeeper that shows AI only the tools you allow

Choose what to expose, tool by tool

A generic tunnel exposes a port. WireCanal exposes tools. Even if your in-house MCP server has 20 tools, the outside AI only sees the ones you have allowed. Tools you have not allowed cannot be called — and they do not even appear in the tool listing (tools/list). To the outside AI, hidden tools simply do not exist.

• check_circleThe default is deny-all (default: deny). Only explicitly allowed tools get through.
• check_circleDestructive operations such as delete and drop can be denied in one stroke (deny_destructive).
• check_circleDenied tools can be hidden from the listing (hide_denied_in_list).

The allowlist stays in your hands (two-key)

The allowlist that decides which tools may be shown to the outside is a configuration file on your own machine — wirecanal.json — and that file is the single source of truth. The cloud side (the dashboard and public servers) cannot rewrite this ledger. Even if something went wrong on the cloud side, no tool would ever be added without your organization's consent — this two-key structure is what makes WireCanal pass corporate approval processes.

The wirecanal.json for an MCP canal adds mode and tools to the usual 2 keys, for 4 keys in total:

{
  "access_key": "ck_...",
  "forward_target": "127.0.0.1:9904",
  "mode": "mcp",
  "tools": { "default": "deny", "allow": ["tool_name_1", "tool_name_2"] }
}

This is not a promise on paper — it is a mechanism that actually runs. If an AI tries to call a tool you have not allowed, the local Agent refuses it on the spot, and the call never reaches your in-house MCP server. We have verified this behavior in real connections with actual AI services.

Trace who came, when, and from where — after the fact

Incoming access is recorded per canal (who, when, from where, and the result). You can retrieve the records from the dashboard, or via the API and CSV. The audit records stay within your organization — WireCanal's cloud side never holds tool arguments or execution results. (When you use a cloud AI, the requests and execution results needed for tool calls are exchanged with the AI service you use. Per-tool behavioral auditing and tamper detection are planned as future work.)

Just pick your AI. OAuth is set up automatically

Answer "Which AI will you use?" in the canal creation wizard, and the remote MCP server URL (e.g. https://<hostname>/mcp) and the connection settings for that AI are prepared automatically. For Claude and ChatGPT, you connect by simply pasting the canal URL (client registration and the choice of authentication method are automatic). For Bestllam and Grok, an OAuth client (ID and secret) for the connection is issued automatically — no need to run your own OAuth server.

• check_circleBestllam: OAuth supported. Paste the URL and client credentials into the connector settings.
• check_circleClaude (claude.ai): just paste the URL into a custom connector (OAuth, auto-configured).
• check_circleChatGPT: just paste the URL into a developer-mode connector (OAuth, auto-configured).
• check_circleGrok (grok.com): register as a custom connector (OAuth).
• check_circleOther AI (Claude Code, Gemini CLI, etc.): issue a connection key and pair it up.

Every time it connects, the AI must pass through WireCanal's connection-consent screen. Only the canal owner can grant approval. The proof of connection is renewed automatically every hour, and revoking the client stops it on the spot.