Your workspace as an MCP server

MCP (Model Context Protocol) is the open standard AI tools use to discover and call tools. It is the common language that opens the whole AI ecosystem inside Runwork: every AI tool your team uses speaks it, and so does everything your team has. It runs in every direction: your workspace serves everything out over MCP, outside MCP servers plug in, and every app your team makes becomes a server of its own.

The workspace MCP server: everything your team has

Every workspace gets a built-in MCP server. One connection exposes all of Runwork's functionality: entities (customers, orders, products), endpoints, workflows, schedules, integrations, components, agents, and file storage. And, just as important, it exposes your team's AI work itself: the skills your people have taught their AI, the apps their agents have made, the workflows that keep running for the team. All mapped to MCP tools with proper schemas and descriptions. Connect Claude Desktop and ask it to "query all customers who signed up this month" or "trigger the weekly report workflow."

Every app your team makes is an MCP server

This is automatic. You don't configure anything. The moment an app your agent made is deployed, it gets its own MCP server URL and its own auto-generated SKILL.md, so the rest of the team's AI tools can use it two ways: call it as an MCP server, or read it as a skill. Add an entity to your app? It's immediately queryable via MCP. Add a workflow? It's triggerable. Connect an integration? It's callable. Every feature you add to your app is instantly available as an MCP tool.

Per-app MCP servers give focused access. Connect Cursor to just your CRM app and it gets CRM tools without seeing your entire workspace. Connect Claude Desktop to your Inventory Tracker and it can query stock levels, trigger reorder workflows, and download reports. Each app's MCP server exposes that app's entities, endpoints, workflows, schedules, integrations, file storage, and agents. Full read and write access.

This is what makes Runwork's AI ecosystem compound. Your team creates 10 apps. That's 10 MCP servers, each with its own tools, plus the workspace server that aggregates everything. Any AI tool your team uses, Claude Code, Cursor, Codex, Claude Desktop, gets access to all of it. The more apps you create, the more your AI tools can do.

Bring outside MCP servers in, synced to everyone

Your apps are only part of it. Connected integrations also appear as MCP tools. Connect HubSpot, and your AI agents get a hubspot_api tool. Connect Stripe, and they can query payments and subscriptions. These integration tools show up alongside your app tools in the MCP Servers Hub.

It works in the other direction too. Paste any MCP server's URL in workspace settings and Runwork connects, authenticates, and discovers its tools, resources, and prompts. From there it syncs to the whole team: after runwork sync, every teammate's MCP clients, which are, generally, their AI tools, can call it. Bring one in from the public MCP Registry, or connect your company's own internal servers: a documentation server, a metrics server, an HR system.

External tools even become skills: each connected server generates skill descriptions, so your agents browse and load its tools the same way they load team-authored skills. It is all allowlisted: every external domain must be approved in workspace settings before an agent can call it, with rate limiting built in. The Hub shows live health status for each server: connected, slow, unreachable, or error, with latency measurements. The Tool Explorer lets you browse and test any tool from any server directly in the dashboard.

Authentication and protocol

Authentication uses your existing workspace API keys. Keys can be scoped to specific apps for per-app MCP access. The same security model you use for your public APIs applies to MCP connections. Both Streamable HTTP and SSE transports are supported. Any MCP-compliant client works with no extra setup.

MCP also exposes resources (entity schemas, app manifests) and prompts (app skills). External tools get tool calls plus the full context of your business behind them. The same skills your teammates' agents use are available to every connected tool too.

MCP during local development

MCP servers connect naturally to local development. When you develop with the Runwork CLI, your local AI tools can connect to your app's MCP server while you code. Claude Code reads SKILL.md (auto-generated for every app, also downloadable from the dashboard) and CLAUDE.md + AGENTS.md (created on the fly by the platform during runwork dev) for framework knowledge, and connects via MCP for live data access. This means your AI assistant can query real data, test workflows, and verify integrations while building features.

The MCP Servers Hub is a dedicated dashboard for managing all MCP connections. Four tabs organize your servers: All, App MCPs, External Servers, and Integrations. Each external server shows live health status with connection latency, tool count, and color-coded indicators (connected, slow, unreachable, error). The Tool Explorer lets you browse every tool exposed by any server and test them directly with parameter inputs and live results. Quick-connect shows copyable endpoint URLs with inline API key creation, so connecting a new AI tool takes seconds. Connected integrations appear as MCP tools too, so your HubSpot, Stripe, and Slack connections are browsable alongside app-specific tools.