WORKING MACHINES GUIDE
What is MCP in AI? A practical Model Context Protocol guide
PUBLISHED AUGUST 29, 2026 / REVIEWED SEPTEMBER 2, 2026
In AI, MCP means Model Context Protocol: a standard way for an AI client to discover and call external tools. A remote MCP integration can expose a small discovery surface while keeping provider credentials in a separate runtime. The agent receives action schemas and results, but OAuth tokens and API keys do not need to be inserted into prompts or stored in the agent process.
What an MCP connection contains
An MCP client connects to a server, negotiates available capabilities, and invokes tools using structured inputs. For an application layer, the visible tools should help the agent search the catalog, inspect an action, choose a connection, and execute. This avoids sending thousands of unrelated schemas to the model.
The MCP authorization flow authenticates the person or workspace using the agent. Provider authorization is a separate concern: each connected application still needs its own account identity and scopes.
- MCP client identity
- Working Machines authorization
- Provider account connections
- Action discovery and schemas
- Structured execution results
Why the endpoint returns an authorization error in a browser
A remote MCP URL is a protocol endpoint, not a normal web page. Opening it directly does not include the authorization header and protocol request expected by the server, so an error such as missing authorization is correct behavior.
A landing-page Connect button should copy the endpoint or show client configuration instead of navigating directly to it. The MCP client then starts the supported authorization flow.
- Copy the server URL from the product page.
- Add it to the MCP client's configuration.
- Start the connection from the client.
- Complete browser authentication and consent.
- Return to the client and verify tool discovery.
Context-efficient tool discovery
Large catalogs create a context problem if every schema is loaded in advance. A discovery-first server exposes a compact set of meta-tools. The agent describes the outcome, receives candidate actions, inspects the exact contract, and loads only what it intends to call.
This pattern improves relevance and makes model context easier to audit. It also lets the catalog grow without linearly increasing every conversation's tool payload.
- Search by outcome and provider
- Rank a bounded action set
- Inspect one action guide
- Select an authorized connection
- Execute the exact action contract
Security boundaries
MCP standardizes communication; it does not make every tool safe automatically. Validate inputs, enforce provider and action policies outside the prompt, guard outbound network access, and treat tool output as untrusted data.
Credentials can remain outside model context while returned business data is still visible to the model. Organizations must align the chosen model's privacy and retention settings with the sensitivity of each workflow.
- Never return provider tokens in tool results
- Use explicit scopes and named connections
- Require confirmation for high-impact writes
- Defend against prompt injection in external content
- Record and review execution outcomes
Primary references
The protocol and security guidance in this article is grounded in the following primary documentation.