connecta

connecta

One MCP to rule them all. A single MCP endpoint that aggregates many downstream connectors — remote MCP servers and plain HTTP APIs — and presents agents a fixed set of nine meta-tools.

Category
Visit Server

README

connecta

A monochrome clay Connecta hub joining many tools

One MCP to rule them all. A single MCP endpoint that aggregates many downstream connectors — remote MCP servers and plain HTTP APIs — and presents agents a fixed set of nine meta-tools instead of hundreds of individual tools.

                                        ┌── remoteMcp("notion")   → mcp.notion.com
Claude / Cursor ── MCP ──▶  connecta ───┼── remoteMcp("linear")   → mcp.linear.app
   sees 9 tools            /mcp         ├── api("resend")         → fetch(...)
                                        └── api("internal")       → fetch(...)

The agent always sees exactly nine tools; it searches when it needs something and describes only what it will call. Progressive disclosure is the point. Connectors are declared in TypeScript (config as code — adding one is a code change + deploy, no database of integrations), and the same fetch-based core runs on Node and Cloudflare Workers. Optionally, code mode adds a tenth tool that runs model-written orchestration code in a sandbox.

Installation

Node deployments require Node.js 20.9 or newer.

npm install @zackbart/connecta

Provider and runtime integrations remain consumer-owned. Install only the optional packages a deployment uses—for example:

npm install @clerk/backend          # optional Clerk auth adapter
npm install quickjs-emscripten      # optional Node code-mode executor
npm install @cloudflare/codemode    # optional Worker code-mode executor

The nine meta-tools

A tool address is <connectorId>.<toolName> (e.g. notion.search).

Tool Input Returns
list_connectors { probe? } live (probe: true, default) or cached health, tool count, and recent real-call observations
skills { name? } lists or fetches the concise usage guide for choosing among the meta-tools
search_tools { query?, connector?, limit?, offset?, fullDescriptions?, includeSchemas? } ranked, paginated matches; optionally includes compact/raw schemas to remove a round trip
describe_tools { addresses[], format?, fullDescriptions? } names, descriptions, input/output schemas, and behavior annotations
call_tool { address, args?, fields?, resultMode?, timeoutMs?, maxRetries?, diagnostics? } invokes only tools explicitly annotated readOnlyHint: true
call_destructive_tool same as call_tool invokes unannotated, write-capable, or destructive tools through a host-visible approval boundary
authorize_connector { connector, force? } starts (or with force, restarts) the downstream OAuth flow; returns the authorizationUrl to open
get_result { id, offset?, maxBytes? } a byte-slice page of a truncated result — { text, offset, nextOffset?, totalBytes }
batch_call { calls, resultMode?, timeoutMs?, maxRetries? } 1–10 parallel calls sharing request-scoped clients, with attempts/timing/errors
execute_code (optional) { code } result + logs of bounded async JS orchestration over explicitly read-only tools; registered only when an executor is configured

Package vs. deployments

This repository owns the reusable package. Each deployment should be a small, separate Worker project that pins an exact package version and owns only its connector configuration, auth policy, domain, bindings, migrations, and secrets. examples/worker/ is the starting template.

Quickstart — Cloudflare Worker

import { DynamicWorkerExecutor } from "@cloudflare/codemode";
import {
  api, bearerToken, createConnecta, remoteMcp,
} from "@zackbart/connecta";
import { clerkAuth } from "@zackbart/connecta/auth/clerk";
import { cloudflareKvStorage } from "./cloudflare-kv.js";

const build = (env: Env) =>
  createConnecta({
    publicUrl: env.PUBLIC_URL,
    storage: cloudflareKvStorage(env.CONNECTA_KV),
    // Code mode (optional): needs a `worker_loaders` binding in wrangler.jsonc.
    executor: new DynamicWorkerExecutor({ loader: env.LOADER }),
    auth: [
      bearerToken(env.CONNECTA_TOKEN),
      clerkAuth({
        publishableKey: env.CLERK_PUBLISHABLE_KEY,
        secretKey: env.CLERK_SECRET_KEY,
        publicUrl: env.PUBLIC_URL,
      }),
    ],
    connectors: [
      remoteMcp("notion", {
        url: "https://mcp.notion.com/mcp",
        auth: { type: "headers", headers: { Authorization: `Bearer ${env.NOTION_TOKEN}` } },
      }),
    ],
  });

// Lazy per-isolate singleton: keeps only serializable tool/catalog data warm.
let connecta: ReturnType<typeof build> | undefined;

export default {
  fetch(request: Request, env: Env): Promise<Response> {
    connecta ??= build(env);
    return connecta.fetch(request);
  },
};

Deployable example: examples/worker/. Its Cloudflare KV and D1 implementations are deployment-owned examples over the generic KVStorage and ActivityStore contracts; they are not package exports.

clerkAuth is an optional adapter. Install @clerk/backend and import it from @zackbart/connecta/auth/clerk only in deployments that use Clerk. Other identity providers can implement the exported InboundAuth interface.

Quickstart — Node

import { api, bearerToken, createConnecta } from "@zackbart/connecta";
import { fileStorage, listen } from "@zackbart/connecta/node";
import { quickJsExecutor } from "@zackbart/connecta/quickjs";

const connecta = createConnecta({
  storage: fileStorage("./.connecta-state.json"), // or memoryStorage()
  auth: bearerToken(process.env.CONNECTA_TOKEN!),
  executor: quickJsExecutor(), // code mode (optional): QuickJS/WASM sandbox
  connectors: [
    api("time", {
      description: "Time — clock utilities",
      tools: [{
        name: "get_now",
        description: "Return the current ISO timestamp.",
        inputSchema: { type: "object", properties: {} },
        annotations: { readOnlyHint: true },
        handler: async () => ({ now: new Date().toISOString() }),
      }],
    }),
  ],
});

listen(connecta, 8787); // http://localhost:8787/mcp

Example: examples/node/. Docker (single-service compose stack): examples/docker/.

Code mode

With an executor configured, the model can write an async function instead of making one call_tool round trip per step — loops, joins across connectors, filtering big responses down in-sandbox before they hit the context window:

async () => {
  const pages = await notion.search({ query: "roadmap" });
  return pages.results.map((p) => p.title);
}

The sandbox has no network/filesystem/env — only explicitly read-only connector globals and connecta.call, connecta.batch, connecta.search, and connecta.describe; credentials stay host-side. A run may make at most 20 host calls, a connecta.batch may contain at most 10 calls, and every host call has a 15-second deadline. Executors: DynamicWorkerExecutor (@cloudflare/codemode, Workers, Worker Loader binding — open beta) or quickJsExecutor() (@zackbart/connecta/quickjs, QuickJS-in-WASM, runs anywhere). Omit the executor and connecta is exactly the nine-tool server. Details: docs §13.

Credentials in /ui

API connectors can declare one operator-managed credential or a named set of credential fields. Connecta renders Add / Replace / Test / Remove controls inside that connector's /ui card, encrypts the values with AES-GCM, and stores only ciphertext in the deployment's existing KVStorage. Credentials are available only to the connector through ctx.credential.get(), get(name), or getAll(); they are never returned by /ui, MCP tools, or code mode.

api("example", {
  description: "Example — authenticated API",
  credential: {
    label: "API token",
    description: "Token used for outbound Example API requests.",
  },
  tools: [{
    name: "get_profile",
    description: "Get the authenticated Example profile.",
    inputSchema: { type: "object", properties: {} },
    annotations: { readOnlyHint: true },
    handler: async (_args, ctx) => {
      const token = await ctx.credential?.get();
      if (!token) throw new Error("Example API token is not configured.");
      return fetch("https://api.example.com/profile", {
        headers: { Authorization: `Bearer ${token}` },
      }).then((response) => response.json());
    },
  }],
})

Set credentialEncryptionKey on createConnecta to a base64-encoded 32-byte key held in the runtime's secret store (openssl rand -base64 32). Credential mutation routes require the configured Clerk provider and a same-origin browser request; the static inbound bearer cannot administer the vault.

Connecta does not bundle service-specific HTTP API connectors. Package consumers define them with api() (or implement Connector directly), keeping endpoint, credential, and tool choices in the consuming project.

Payload-free tool activity

Connecta can record which resolved downstream tools were actually invoked without storing their arguments, results, generated code, search text, or raw errors. Supply a vendor-neutral activity store:

const events: ToolCallActivityEvent[] = [];

const connecta = createConnecta({
  connectors,
  activity: {
    record(event) {
      events.push(event);
    },
    async list({ limit }) {
      return { events: events.slice(-limit).reverse() };
    },
  },
});

One final event is emitted for each resolved connector call made through call_tool, batch_call, or execute_code; retries remain one event with an attempts count. Implementing list enables the authenticated Activity tab in /ui; activityReadGate can narrow reads further. Writes are best-effort and never change a tool result. Clerk calls carry the Clerk user ID; shared bearer calls are honestly labeled as bearer unless bearerToken(secret, { subjectId }) assigns that credential a stable subject.

Learn more

  • docs/documentation.md — how everything works: architecture, the meta-tools, connectors, inbound auth, downstream OAuth, storage, running it (Node / Workers / Docker), Clerk setup, testing, troubleshooting.
  • docs/design.md — why it's built this way, and the non-goals.
  • Status UI — a read-only operator dashboard at GET /ui (open shell; data from the bearer/Clerk-gated /ui/data). See §14.

Recommended Servers

playwright-mcp

playwright-mcp

A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.

Official
Featured
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.

Official
Featured
Local
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.

Official
Featured
Local
TypeScript
VeyraX MCP

VeyraX MCP

Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.

Official
Featured
Local
graphlit-mcp-server

graphlit-mcp-server

The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.

Official
Featured
TypeScript
Kagi MCP Server

Kagi MCP Server

An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.

Official
Featured
Python
Neon Database

Neon Database

MCP server for interacting with Neon Management API and databases

Official
Featured
E2B

E2B

Using MCP to run code via e2b.

Official
Featured
Exa Search

Exa Search

A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.

Official
Featured
Qdrant Server

Qdrant Server

This repository is an example of how to create a MCP server for Qdrant, a vector search engine.

Official
Featured