Documentation
¶
Overview ¶
Package mcp exposes the coreutils tool registry over the Model Context Protocol — the third consumption surface (alongside the in-process shell ExecHandler and the busybox multicall binary) so that non-Go agents (codex, claude, …) can drive the AgentOS userland.
The registry holds CLI-shaped tools (argv + stdin -> stdout/stderr/exit), so the faithful MCP mapping is two generic meta-tools rather than a schema per tool:
- list_tools — enumerate the tools this build ships (name + synopsis)
- run_tool — run one: {name, args, stdin, dir, env} -> {stdout, stderr, exit_code}
run_tool delegates to multicall.Dispatch, so name resolution and the unknown-tool diagnostic match the standalone `coreutils` binary exactly.
As agentic verbs (symbols, repomap, …) land in the registry they can additionally be registered as typed, individually-schema'd MCP tools on the same server via RegisterTool; the generic pair always covers the whole registry as a floor.
This package uses the official SDK (github.com/modelcontextprotocol/go-sdk), the same one the umbrella already pins, aliased mcpsdk to avoid clashing with this package's own name.
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func NewServer ¶
NewServer builds an MCP server exposing the current tool registry via the generic list_tools / run_tool pair. Callers must have blank-imported the tool sets they want available (e.g. coreutils/cmds/all) before any client lists or runs tools.
func ServeStdio ¶
ServeStdio runs an MCP server over stdio until the transport closes or ctx is cancelled — the entrypoint a `coreutils mcp` / `bashy mcp` front end calls. A client closing stdin (io.EOF) is the normal way to end a stdio session, so it is reported as a clean shutdown (nil), not an error.
Types ¶
type ListToolsInput ¶
type ListToolsInput struct{}
ListToolsInput is empty — list_tools takes no arguments.
type ListToolsOutput ¶
type ListToolsOutput struct {
Tools []ToolInfo `json:"tools" jsonschema:"every tool this build ships, sorted by name"`
}
ListToolsOutput is the list_tools result.
type RunToolInput ¶
type RunToolInput struct {
Name string `json:"name" jsonschema:"tool to run, e.g. \"grep\""`
Args []string `json:"args,omitempty" jsonschema:"argument vector after the tool name"`
Stdin string `json:"stdin,omitempty" jsonschema:"standard input fed to the tool"`
Dir string `json:"dir,omitempty" jsonschema:"working directory; relative operands resolve against it"`
Env map[string]string `json:"env,omitempty" jsonschema:"environment for the invocation; empty means none"`
}
RunToolInput is the run_tool argument set.
type RunToolOutput ¶
type RunToolOutput struct {
Stdout string `json:"stdout" jsonschema:"captured standard output"`
Stderr string `json:"stderr" jsonschema:"captured standard error"`
ExitCode int `json:"exit_code" jsonschema:"process exit status (0 success, 2 usage error, …)"`
}
RunToolOutput is the run_tool result.
type ToolInfo ¶
type ToolInfo struct {
Name string `json:"name" jsonschema:"command name as spelled on the CLI"`
Synopsis string `json:"synopsis" jsonschema:"one-line description"`
Usage string `json:"usage" jsonschema:"usage line"`
Group string `json:"group,omitempty" jsonschema:"Command Atlas functional group (fileutils, textutils, code-intel, …)"`
Caps []string `json:"caps,omitempty" jsonschema:"Command Atlas capability flags (json, read-only, destructive, …)"`
}
ToolInfo describes one registered tool for list_tools. Group and Caps are the Command Atlas axes (pkg/atlas): the functional group and the agentic capability flags; both are additive and omitted when the atlas has no entry for the name.
Source Files