MCP Tools

Agents can consume any remote MCP server that speaks the streamable HTTP transport — the same protocol almyty itself serves. Point almyty at the server once, and every tool it exposes becomes a regular almyty tool: usable in agent pipelines, servable through gateways, and testable in the tool tester.

Adding an MCP server

On the Tools page, click Add MCP Server and enter a name, the server URL, and an optional bearer token (stored encrypted, never returned by the API). almyty performs the MCP handshake (initialize → tools/list), and registers each remote tool with an MCP badge.

Re-sync a source at any time to pick up added, changed, or removed tools — removed remote tools are marked inactive so run history and gateway assignments survive. Deleting a source removes its tools.

Execution

When an agent calls an MCP tool, almyty issues tools/call against the source with the agent’s arguments and maps the result back as the tool output. Remote failures surface as typed tool errors on the run (MCP_CONNECT_FAILED, MCP_TIMEOUT, MCP_REMOTE_ERROR, …) — never an opaque crash.

Security

Server URLs are validated before every request: private ranges, loopback, link-local, and cloud-metadata addresses are blocked, and redirects are refused. Self-hosted deployments that need in-cluster MCP servers can set MCP_ALLOW_PRIVATE_URLS=true (http/https only).
Auth secrets (bearer tokens, custom headers) are AES-256-GCM encrypted at rest.

Supported protocol

Streamable HTTP transport, protocol revision 2025-03-26 and later (both plain-JSON and SSE response framings; Mcp-Session-Id handled automatically). Not supported: stdio servers, the legacy HTTP+SSE transport, and server-initiated sampling — expose those through a bridge if you need them.