Mcpproxy Go

MCPProxy solves a problem that arrives quietly. You add a filesystem server, then GitHub, then Playwright, then Jira, then Slack and PostgreSQL, and suddenly every one of your AI clients is loading hundreds of tool definitions before it does any work. MCPProxy puts one endpoint in front of all of them and indexes the tools with BM25, so the agent asks for a capability and gets back only the relevant schemas. Per the docs, that saving only applies in the retrieve_tools and code_execution routing modes; direct routing still lists upstream tools individually. The access model is the part that separates this from a plain connection multiplexer. Profiles group servers into explicit scopes, and scoped credentials let you pin a specific client or agent to a profile, so your research agent doesn't inherit your production database tools. v0.70.0 pushed further with per-tool policy inside profiles, per-client credentials, a single needs-attention list and agent server discovery — a meaningful step up from the seed data's description of profiles plus tokens alone. On security, MCPProxy is honest in a way that matters. New servers are quarantined until approved. Tool Poisoning Attack patterns are scanned for, tool definition changes are held for review under the default manual trust mode, sensitive data in arguments and responses is flagged, and schema checks run against definitions. The docs also spell out the exceptions rather than hiding them: input-schema-only edits outside scan mode, a newly added output schema, and a description added to a tool approved with an empty description all slip past manual review, and scan or auto modes admit more without a human. If you want automatic safety, this isn't it — and the project says so. Where it fits: developers and small teams on macOS, Windows or Linux who already run MCP servers locally and want to stop duplicating configuration across clients. The CLI checks required tools and orchestrates calls in JavaScript or TypeScript, which makes it usable in repeatable workflows rather than just interactive use. Where it doesn't: anyone who wants a point-and-click setup, teams needing cross-machine hosted access with role-based permissions, or buyers who want a vendor-managed cloud proxy. The core server binds to 127.0.0.1:8080 by default, LAN access is optional, and the proxy has no AI model of its own — it depends on upstream servers being reachable.