Local open-source MCP gateway that lets multiple AI clients share tool servers while keeping credentials local.
Toolport is a local-first MCP gateway that lets multiple AI clients share one registry of MCP servers instead of configuring each server separately in Claude, Cursor, Codex, VS Code, Windsurf and other clients 12. Its main architectural idea is lazy discovery: clients load a small set of gateway meta-tools and search for downstream tools when needed rather than placing every tool definition in context on every request 1. Toolport also centralizes agent rules and permissions, supports per-tool disabling, records tool-call activity, fingerprints tool definitions for changes, and can quarantine suspicious tool-description changes 12. The local app and gateway are MIT-licensed and run on Windows, macOS and Linux 12.
The local Toolport app and gateway are free and open source with no account or cloud dependency 12. Toolport Teams is a separate open-core offering: the current repository documentation describes a free tier for up to five people and a paid Team tier at $39 per month for up to five people plus $12 per additional person, with extra access controls, roles, budgets, full audit history and team alerts 2. Teams can use the hosted service or self-host the team control plane, while each member's MCP credentials remain local 2. This gives individual developers a strong value proposition; team economics are still modest relative to enterprise developer tooling.
Toolport's security posture is unusually explicit for a small developer utility. Secrets are stored in the operating-system keychain, the local app requires no Toolport account or cloud service, destructive-tool policies can be enforced across clients, and the repository is MIT-licensed for inspection 12. Official material also publishes reproducible benchmark methodology for its token-savings claims, but those measurements remain first-party and should not be treated as independent performance validation 12. The project documents an upstream Linux WebKit/GLib soundness issue affecting one binding path and explicitly states that it is not remote code execution and does not affect Windows or macOS builds 2. Two independent current product overviews corroborate the local-first architecture, multi-client sharing, keychain storage and governance feature set, but neither provides a substantive reliability benchmark or user-rating corpus 34. The canonical 50-source panel likewise produced no eligible exact-product numeric observation.
Toolport is most useful for developers who use several agent clients and have enough MCP servers that duplicated configuration and tool-definition context become noticeable operational costs. It is also attractive to security-conscious developers who prefer credentials to remain in the OS keychain and want one policy layer for destructive tools across clients 12. The Open WebUI integration demonstrates that the gateway can expose a local HTTP/OpenAPI surface as well as client-specific MCP configuration 5. Toolport's own documentation cautions that small local models around the 7B class may struggle with the lazy search-then-call flow; in that case users can expose the full tool catalog instead 5.
The strongest combination is interoperability, local control and open-source transparency. One registry can feed many AI clients, keys stay local, and governance is applied centrally rather than copied into each client 12. Lazy discovery plausibly addresses a real context-cost problem, and the project publishes code and benchmark material that can be inspected or reproduced, but AiToolMap does not treat the advertised savings figures as independently validated 12. The main weakness is maturity: the project has limited independent testing, no substantial current user-rating corpus and a smaller operational track record than established developer-infrastructure products 34. Users should also expect some client-specific integration differences and platform-edge cases as MCP clients continue changing their configuration models 25.
Not separately stated in the source review.
Products matched by shared tasks, audience, workflow, product type and capabilities.