Skip to content

Part II · Instance setup

MCP Servers

MCP servers are the tool backends of Foundry — the deliberate sibling of Providers, which are the model backends. Providers give agents something to think with; MCP servers give them something to act with: editing artifacts, driving a browser, querying the design system, reading decision records. On a typical instance the tab shows on the order of 10 servers · 84 tools.

Instance Settings · MCP Servers
The MCP Servers tab listing built-in loopback presets and external servers with transports, tool counts, and Test actions
Live rigThe tool registry: built-in loopback presets alongside external servers like Prism and Context7, each with its transport, discovered tool count, and enable/disable control.

Built-in presets and external servers

The registry mixes two kinds of entries:

  • Built-in presets — loopback servers Foundry ships with, covering its own tool surfaces: Artifact editing, OpenCode tools, Browser / Chrome DevTools, Design system, Execution / orchestration, Mock build & screenshot, Decision records, and Vision. You can configure and disable presets, but not delete them — agents and workflows depend on their existence, so removal is a code-level decision, not a settings click. Disabling one is honored everywhere: its tools disappear from every picker and from the runtime.
  • External servers — real network endpoints an admin registers, such as Prism (code intelligence) or Context7 (library docs). These carry an endpoint, credentials, and a Test button.

Transports

A server speaks one of three transports: remote-http (a network MCP endpoint), local-stdio (a locally spawned process), or loopback (Foundry's own in-process tool surfaces — the presets).

Authentication

External servers support four auth modes:

  • None
  • Bearer — a static token, encrypted at rest with the same crypto as provider credentials.
  • Custom headers — for backends with non-standard header schemes.
  • OAuth 2.1 — the full flow for servers that require it.
MCP server · OAuth setup
OAuth configuration in the server drawer with optional client ID, sealed secret, prefilled redirect URI, and Connect action
Live rigThe OAuth 2.1 drawer: client ID left empty triggers dynamic client registration, the redirect URI is prefilled, and Connect… runs the browser flow with PKCE.

OAuth 2.1, in practice

The OAuth drawer is built so that the common case needs almost nothing typed in:

  • Client ID is optional. Leave it empty and Foundry uses dynamic client registration (RFC 7591) — the server issues a client on the spot. Fill it only when the server operator gave you a pre-registered client.
  • The client secret is sealed once saved: write-only, never displayed again.
  • The redirect URI is prefilled — it must match what Foundry's callback actually is, so it is not a field you should have to compose by hand.
  • Connect… starts the browser flow (with PKCE) and stores the resulting tokens, refreshing them automatically thereafter.

Test & Save, and where tools go

Test & Save performs a real MCP tools/list against the server — proving connectivity, auth, and protocol in one step — and records the discovered tools in the registry. From there, discovered tools become selectable per agent: which agent may use which tools is decided in the agent's configuration (the Tools & Subagents tab of the Agent Editor), not globally. The registry answers "what exists and does it work"; the per-agent grant answers "who may use it".

A server can be enabled and still unavailable

Some presets are environment-gated: Prism, for instance, needs its environment configured on the host. Such a server shows as unavailable with a note explaining why, and its tools are greyed out in pickers rather than being offered to agents that could never call them.