mcp
MCP components: connecting tools and services to an agent.
mcp
An mcp describes an MCP server: its source, how it starts, its compatibility
with a harness, the permissions it expects, and the limits on its use. Through
MCP an agent gets structured access to a tool surface: files, a browser, a SaaS
API, databases or local helpers.
ai_stp holds the configuration and the checks. Secrets, tokens and passwords
do not go into a passport.
When you need an MCP
| Task | Does it need mcp? | Note |
|---|---|---|
| The agent must read GitHub issues | yes | MCP defines the connection and the tool surface |
| The agent must know the code review rules | no | that is an instruction or a skill |
| The agent must call a local scanner | possibly | if the scanner is offered as an MCP server |
| An endpoint and access mode must be recorded | yes | but without the token value |
| A command must run from a shortcut | no | that is a command |
How ai_stp reduces the risk
| Risk | The mechanism |
|---|---|
| Secrets in the config | the passport holds no secret values |
| Unexpected permissions | required permissions and warnings are shown |
| A substituted source | the version pins the source, the commit and path, or the package |
| The wrong harness | compatibility is checked before the provider plan |
| An unsafe install | unknown objects need explicit consent |
!!! warning "MCP widens what an agent can reach" Even a good MCP server can give an agent access to data or operations it did not have before. Install it as a tool with permissions, not as "one more Markdown file".
What the user sees
=== "In the catalog" What the MCP is for, the author, the trust line, the supported harnesses and the required permissions.
=== "In the install plan" Which entries the provider will add or change, which secrets must be supplied through the environment or the system store, and how to roll back.
=== "After applying"
status must show the active MCP, the source of the version, and the
result of the provider's check.