mcptools

package
v0.2.0 Latest Latest
Warning

This package is not in the latest version of its module.

 Go to latest Published: Oct 12, 2026 License: Apache-2.0 Imports: 23 Imported by: 0

Documentation

Overview

Package mcptools builds the MCP tool server every levelrail-mcp transport (stdio, network, and the in-process transport an embedded caller like internal/ai uses) shares: one *mcp.Server with every tool registered against the same apiclient.Client. Pulled out of cmd/levelrail-mcp (which cannot be imported, being package main) so a non-CLI caller can build the identical tool server without spawning a subprocess.

Index

Constants

View Source 
const (

	// MetaSensitiveKey is the tool _meta key set on sensitive tools.
	MetaSensitiveKey = "levelrail/sensitive"
	// MetaUntrustedKey is the tool _meta key set on tools whose result carries untrusted text.
	MetaUntrustedKey = "levelrail/untrusted-output"
)
View Source 
const (
	// ModeReadOnly registers read tools only.
	ModeReadOnly Mode = "read-only"
	// ModeStandard registers read and mutating tools, not destructive ones.
	ModeStandard Mode = "standard"
	// ModeFull registers every tool.
	ModeFull Mode = "full"

	// EnvMode names the env var that selects the Mode.
	EnvMode = "APP_MCP_MODE"
	// EnvToolsets names the env var that selects the toolsets.
	EnvToolsets = "APP_MCP_TOOLSETS"
	// EnvAllowWrite names the env var that exposes the confirm-gated write tools.
	EnvAllowWrite = "APP_MCP_ALLOW_WRITE"
	// EnvOutputSchemas names the env var that keeps outputSchema in tools/list
	// for levelrail-mcp, which omits it by default.
	EnvOutputSchemas = "APP_MCP_OUTPUT_SCHEMAS"
	// EnvToolProfile names the env var that selects the tool profile.
	EnvToolProfile = "APP_MCP_TOOL_PROFILE"

	// ProfileAgentCore is a small allowlist of the tools an agent needs to
	// ship and debug an app.
	ProfileAgentCore = "agent-core"
)
View Source 
const (
	EnvWaitWindow   = "APP_MCP_WAIT_WINDOW_SECONDS"
	EnvWaitInterval = "APP_MCP_WAIT_POLL_SECONDS"
)

Env vars bounding one wait_for_deploy call, in seconds.

View Source 
const EnvResourceMaxBytes = "APP_MCP_RESOURCE_MAX_BYTES"

EnvResourceMaxBytes names the env var capping a docs resource's size.

Variables

This section is empty.

Functions

func AgentCoreTools 

func AgentCoreTools() []string

AgentCoreTools returns the sorted agent-core allowlist.

func Groups 

func Groups() []string

Groups returns every toolset name, sorted.

func NewServer 

func NewServer(client *apiclient.Client) *mcp.Server

NewServer builds the MCP server and registers every tool against client. Callers can connect to it directly over an in-memory transport (mcp.NewInMemoryTransports), without spawning a real process or touching stdio.

func ParseToolProfile 

func ParseToolProfile(s string) (string, error)

ParseToolProfile validates a tool profile name; empty means no profile.

func ParseToolsets 

func ParseToolsets(s string) ([]string, error)

ParseToolsets parses a comma separated toolset list; empty means all.

func PlannableTools 

func PlannableTools() []string

PlannableTools returns the sorted names of mutating tools plan_change can plan.

Types

type AppEnvView 

type AppEnvView struct {
	Name          string            `json:"name"`
	Env           map[string]string `json:"env"`
	SensitiveKeys []string          `json:"sensitive_keys"`
	SecretKeys    []string          `json:"secret_keys"`
}

AppEnvView is an app's visible plain env vars, plus the names, never values, of its secrets and of its sensitive plain vars.

type Class 

type Class int

Class is a tool's blast radius, the single source for its MCP annotations, the server modes, and the assistant's confirmation gate. 

const (
	// ClassUnset is the zero value of an unclassified tool.
	ClassUnset Class = iota
	// ClassRead tools change nothing.
	ClassRead
	// ClassMutate tools change state but can be undone or re-run.
	ClassMutate
	// ClassDestructive tools delete, roll back or tear down state.
	ClassDestructive
)

func (Class) String 

func (c Class) String() string

type EnvChangeResult 

type EnvChangeResult struct {
	Name           string `json:"name"`
	Key            string `json:"key"`
	Secret         bool   `json:"secret"`
	RedeployNeeded bool   `json:"redeploy_needed"`
	Hint           string `json:"hint"`
}

EnvChangeResult reports which key changed, never any value.

type Meta 

type Meta struct {
	Class Class
	Group string
	// contains filtered or unexported fields
}

Meta is one tool's classification.

func Lookup 

func Lookup(name string) (Meta, bool)

Lookup returns a tool's classification.

func (Meta) Outbound 

func (m Meta) Outbound() bool

Outbound reports whether the tool reaches outside the control plane.

func (Meta) Sensitive 

func (m Meta) Sensitive() bool

Sensitive reports whether the tool touches credentials, secrets or command output.

func (Meta) Untrusted 

func (m Meta) Untrusted() bool

Untrusted reports whether the tool's result carries attacker-influenced text.

func (Meta) WriteGated 

func (m Meta) WriteGated() bool

WriteGated reports whether the tool is hidden unless the server runs with AllowWrite.

type Mode 

type Mode string

Mode bounds which tool classes a server registers. The API token's own abilities still apply on top: a mode only shrinks what the model sees.

func ParseMode 

func ParseMode(s string) (Mode, error)

ParseMode parses a mode name, defaulting an empty value to standard.

type Options 

type Options struct {
	Mode     Mode
	Toolsets []string
	// Profile, when set, replaces the mode's class filter with an allowlist.
	// Read-only mode still hides every non-read tool.
	Profile string
	// AllowWrite exposes the confirm-gated write tools. Off by default.
	AllowWrite bool
	// OmitOutputSchemas drops outputSchema from tools/list. Results still
	// carry structured content; only the schema advertisement is left out.
	OmitOutputSchemas bool
}

Options selects the tools a server exposes.

type PlanChange 

type PlanChange struct {
	Field  string `json:"field"`
	Before string `json:"before"`
	After  string `json:"after"`
}

PlanChange is one field a planned call would change. Values for secrets and env are never included, only whether they are set.

type PlanResult 

type PlanResult struct {
	Tool      string       `json:"tool"`
	DryRun    bool         `json:"dry_run"`
	Executed  bool         `json:"executed"`
	Plannable bool         `json:"plannable"`
	Summary   string       `json:"summary"`
	Changes   []PlanChange `json:"changes,omitempty"`
	// Blocked is true when a freeze window or a required approval would stop
	// the call from taking effect now.
	Blocked  bool     `json:"blocked"`
	Blockers []string `json:"blockers,omitempty"`
	Note     string   `json:"note,omitempty"`
	// Detail carries a richer plan when one already exists (a resource plan,
	// a promotion preview).
	Detail map[string]any `json:"detail,omitempty"`
}

PlanResult is what plan_change returns: what a mutating tool call would do, computed from reads only. Executed is always false.

type Summary 

type Summary struct {
	Mode     Mode
	Toolsets []string
	Profile  string
	Total    int
	Read     int
	Mutate   int
	Destruct int
}

Summary reports what a server exposes.

func NewServerWithOptions 

func NewServerWithOptions(client *apiclient.Client, opts Options) (*mcp.Server, Summary)

NewServerWithOptions builds the MCP server and then unregisters every tool the mode and toolsets exclude, so a hidden tool costs no model context and cannot be called.

Source Files

View all Source files

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL
go.dev uses cookies from Google to deliver and enhance the quality of its services and to analyze traffic. Learn more.