Open source
MCP is how AI agents call tools. It was designed for a trusted, local connection on one machine. SMCP keeps MCP's tool model exactly as it is and adds what you need once tools run across a network or between agents: authentication, encryption and coordination.
SMCP is a standalone project with its own client, server and agent-to-agent specification, a written wire protocol with conformance vectors, and an automated test suite.
What it adds
Agents authenticate with an API key or JWT session, optionally with asymmetric tokens a client can verify but cannot forge, or through your own identity provider.
Payloads are encrypted and authenticated per message, with key exchange for forward-secret session keys.
Signed messages carry a freshness window and are accepted only once, so a captured message cannot be sent again.
Server and client refuse to start with empty, weak or publicly known secrets, and refuse plaintext to anything but localhost.
Agent-to-agent discovery and orchestration, sequential or in parallel, across nodes that verify each other.
A bridge in both directions: standard MCP clients can call SMCP tools, and MCP tools can be served through SMCP's security layer.
With malgra
malgra is a full SMCP peer, verified to interoperate byte for byte with the reference implementation. Tools served through malgra can be called over SMCP, and SMCP's safety hooks can hand tool calls to malgra's policy engine for injection and tool-poisoning checks.
The gateway underneath eutonomous: rules, personal-data replacement, tool-output scanning and a signed record on every request.
A terminal coding agent that can use MCP and SMCP tool servers, built to take advantage of malgra.
Encrypted long-term memory for agents, served to whole fleets of agents over SMCP, one namespace per API key.