Show HN: Lensa – Open-source MCP connectors and skills for ChatGPT

What happened

The raw HTML is not useful. Let me search for the actual README content in the file. The page returned raw HTML. Let me try to fetch the raw README directly. Now I have solid information from the README. Let me write the analytical piece based only on confirmed facts.

---

The technical detail

A developer going by mitrotasios published Lensa to GitHub, presenting it as an open-source Model Context Protocol server designed to act as a single local gateway through which MCP clients — including ChatGPT and Claude — can call into multiple third-party application APIs. The project is tagged at version 0.1.0, described explicitly as a work in progress intended for local, single-user use only. It surfaced on Hacker News under the Show HN banner, drawing one point and no comments at time of writing, which places it squarely in the category of very early-stage tool that has yet to build an audience.

The core architecture treats connectors and skills as two distinct layers. Connectors are typed, focused wrappers around a specific application's API. Lensa ships two at launch: one for Clerk, covering users, sessions, organizations, memberships, invitations, domains, enterprise connections, and access actions; and one for Vercel, covering projects, deployments and logs, environment variables, domains, DNS, team membership, and deployment actions. Skills sit on top of connectors and describe named reusable workflows — things like a Clerk login investigation flow or a Vercel deployment diagnosis and incident triage sequence. A user enabling only a subset of tools does not expose the full surface area of any given connector to the model.

Why it matters for builders

The gateway runs as a Docker container. Credentials for each enabled connector go into a YAML secrets file, and a separate configuration file — lensa.yaml — governs which connectors, tools, and skills are active. The project notes explicitly that the default checked-in configuration enables all current tools, including write tools that are gated on a confirmation step, and recommends starting from a disabled example configuration and adding only what is needed. Connecting to ChatGPT requires an additional step: obtaining an OpenAI Secure MCP Tunnel ID and runtime key, which suggests it is using OpenAI's remote MCP capability rather than a locally registered tool definition.

For engineers building or maintaining products that use Clerk for auth or Vercel for deployment, the practical value is in the skills layer. Instead of constructing a bespoke prompt chain each time an incident needs investigating — pulling session state, checking deployment logs, inspecting environment variable history — the skill can encode that diagnosis sequence once and replay it through natural language. Because the skills are defined in configuration and the catalog is documented, teams can version-control investigation playbooks the same way they version-control runbooks. That is meaningfully different from an ad-hoc MCP setup where the model improvises tool use from raw API descriptions.

Market and competitive context

Competitively, this occupies the same conceptual space as other MCP aggregation layers but takes a different approach to scope control. Many MCP implementations give the model access to an entire API surface and rely on prompt engineering to prevent destructive operations. Lensa makes scope a configuration concern rather than a prompt concern: if a tool is not listed in lensa.yaml, the model cannot call it. The confirmation-gated write tools add a second guardrail. Projects like Zapier's MCP integration and various community-built connector registries take a broader marketplace approach, optimizing for coverage; Lensa takes the opposite stance and optimizes for auditability and minimal exposure.

Practically, anyone evaluating Lensa should look at the connector authoring documentation before committing. The project explicitly notes that the two current connectors are not a closed catalog and that adding a connector starts with a documented guide. The durability of the project depends on whether that contribution path is low enough friction to attract community-built connectors for services beyond Clerk and Vercel. The skills model is the more novel bet: if teams find that codifying investigation and triage workflows as named skills provides consistent diagnostic value, that pattern could generalize into something like an operational knowledge base that runs directly inside the model's tool call loop.

What to watch next

The main open questions involve the ChatGPT integration specifically. OpenAI's Secure MCP Tunnel is a relatively new mechanism and its stability, latency characteristics, and how it interacts with the server's tool-selection filtering are not documented in the README. There is also no mention of how the server handles concurrent model sessions or whether the local single-user constraint is architectural or just a warning about the current authentication model. Because the project is pre-release and self-hosted, credential management is entirely the deployer's responsibility — storing Clerk and Vercel API keys in a local YAML file is appropriate for a developer laptop but would need a different approach before this pattern could be used in a shared team environment.