MCP Tools - agtOS

agtOS exposes an MCP (Model Context Protocol) server via Streamable HTTP transport. External AI clients can connect to invoke voice pipeline, system monitoring, workflow, and scheduling tools. Endpoint: http://<host>:<MCP_PORT>/mcp Default: http://localhost:4100/mcp Transport: Streamable HTTP (ADR-005) Server Name: agtos | Server Version: 1.0.0

Voice Tools

voice.speak — Speak Text

Synthesize text to speech and play it through the agtOS voice pipeline. Returns confirmation when synthesis begins.

Input Schema

text

string

required

The text to speak (min length: 1).

voice

string

Voice name override (e.g. "af_heart").

sessionId

string

Target voice session ID. Omit for default session.

Example Input

{
  "text": "Good morning. The weather today is sunny with a high of 72 degrees.",
  "voice": "af_heart"
}

Success Response

Speaking: "Good morning. The weather today is sunny with a high of 72 deg..." [status: queued]

Error Response

Error: TTS provider unavailable

voice.listen — Listen for Speech

Start listening for speech input via the agtOS voice pipeline. Returns the transcribed text when speech is detected and processed.

Input Schema

durationMs

number

Maximum listen duration in milliseconds (default: 5000, max: 30000). Must be a positive integer.

sessionId

string

Target voice session ID. Omit for default session.

Example Input

{
  "durationMs": 10000
}

Success Response

Turn on the kitchen lights and set a timer for 5 minutes.

If no speech is detected:

(no speech detected)

Error Response

Error: STT provider unavailable

System Tools

system.health — System Health

Get comprehensive system health status including all registered services (Redis, STT, TTS, Ollama, Claude). Returns per-service status, response times, and overall system health.

Input Schema

No input parameters.

Success Response

{
  "status": "healthy",
  "services": {
    "redis": { "status": "healthy", "responseTime": 2 },
    "stt-speaches": { "status": "healthy", "responseTime": 45 },
    "tts-speaches": { "status": "healthy", "responseTime": 38 },
    "ollama": { "status": "healthy", "responseTime": 12 }
  },
  "metrics": {
    "requestsPerMinute": 15,
    "errorsPerMinute": 0,
    "activeSessions": 2,
    "latency": {
      "p50": 120,
      "p95": 450,
      "p99": 800
    }
  }
}

Error Response

Health check failed: Redis connection refused

session.status — Session Status

Get information about voice sessions. When called without a sessionId, returns a summary of all active sessions. When called with a specific sessionId, returns detailed information about that session.

Input Schema

sessionId

string

Specific session ID to query. Omit for overview of all sessions.

Example Input (all sessions)

{}

Example Input (specific session)

{
  "sessionId": "session-a1b2c3d4e5f6"
}

Success Response (all sessions)

{
  "sessions": [
    { "id": "session-a1b2c3d4e5f6", "isActive": true },
    { "id": "session-f6e5d4c3b2a1", "isActive": true }
  ],
  "activeSessions": 2
}

Error Response

Error: Session not found

Orchestration Tools

Orchestration tools are conditionally registered. Workflow tools require a workflow engine; scheduler tools require a Redis-backed scheduler. If neither is available, orchestration tools are skipped entirely.

workflow.run — Run Workflow

Execute a registered workflow by ID with optional input. Returns the execution ID and final state.

Input Schema

workflowId

string

required

ID of the registered workflow to execute (min: 1).

input

object

Optional input payload to pass to the workflow.

Example Input

{
  "workflowId": "morning-routine",
  "input": {
    "location": "San Francisco"
  }
}

Success Response

{
  "executionId": "exec-789abc",
  "workflowId": "morning-routine",
  "state": "completed",
  "startTime": 1711612800000,
  "endTime": 1711612802000
}

Error Response

Error: Workflow engine not available

Error: Workflow 'nonexistent' not found

workflow.list — List Workflows

List all registered workflow definitions. Returns workflow IDs, names, descriptions, and step counts.

Input Schema

No input parameters.

Success Response

[
  {
    "id": "morning-routine",
    "name": "Morning Routine",
    "description": "Plays morning briefing and checks calendar",
    "stepCount": 3
  }
]

Error Response

Error: Workflow engine not available

schedule.create — Create Scheduled Task

Create a new scheduled task. Supports cron expressions for recurring tasks, one-time execution at a timestamp, or interval-based repetition.

Input Schema

name

string

required

Human-readable name for the scheduled task (min: 1).

scheduleType

string

required

Type of schedule: cron, once, or interval.

expression

string

Cron expression (e.g. "*/5 * * * *"). Required when scheduleType is cron.

atTimestamp

number

Unix millisecond timestamp. Required when scheduleType is once.

intervalMs

number

Interval in milliseconds. Required when scheduleType is interval.

eventTopic

string

required

Event bus topic to publish when the task fires (min: 1).

payload

object

Optional payload included in the fired event.

Example Input (cron)

{
  "name": "Morning briefing",
  "scheduleType": "cron",
  "expression": "0 7 * * *",
  "eventTopic": "briefing.morning",
  "payload": { "workflowId": "morning-routine" }
}

Example Input (once)

{
  "name": "Meeting reminder",
  "scheduleType": "once",
  "atTimestamp": 1711699200000,
  "eventTopic": "reminder.fire",
  "payload": { "message": "Team standup in 5 minutes" }
}

Example Input (interval)

{
  "name": "System heartbeat",
  "scheduleType": "interval",
  "intervalMs": 60000,
  "eventTopic": "system.heartbeat"
}

Success Response

{
  "taskId": "task-abc123",
  "name": "Morning briefing",
  "status": "active",
  "nextRunAt": 1711699200000,
  "scheduleType": "cron"
}

Error Response

Error: Task scheduler not available

Error: "expression" is required for cron schedules

schedule.list — List Scheduled Tasks

List all scheduled tasks with optional status filtering. Returns task IDs, names, statuses, schedule types, and next run times.

Input Schema

status

string

Filter by status: active, paused, completed, cancelled, failed.

Example Input

{
  "status": "active"
}

Success Response

[
  {
    "id": "task-abc123",
    "name": "Morning briefing",
    "status": "active",
    "scheduleType": "cron",
    "nextRunAt": 1711699200000,
    "lastRunAt": 1711612800000,
    "runCount": 5
  }
]

Error Response

Error: Task scheduler not available

schedule.cancel — Cancel Scheduled Task

Cancel a scheduled task by ID. The task will no longer fire.

Input Schema

taskId

string

required

ID of the scheduled task to cancel (min: 1).

Example Input

{
  "taskId": "task-abc123"
}

Success Response

Task task-abc123 cancelled successfully

Error Response

Error: Task scheduler not available

Error: Task not found

Memory Tools

memory.ask_about_user — Ask About User

Ask a question about the user via the Dialectic reasoning engine. Gathers the user profile, stored conclusions, and relevant episodic memories to synthesize an answer with a confidence score.

Input Schema

question

string

required

The question to ask about the user (max 2,000 characters).

Example Input

{
  "question": "What programming languages does this user prefer?"
}

Success Response

{
  "answer": "Based on conversation history, the user primarily works with TypeScript and Python. They prefer TypeScript for backend services and Python for data analysis.",
  "confidence": 0.82,
  "sources": ["ep-abc123", "ep-def456", "ep-ghi789"]
}

Error Response

Error: Memory coordinator not available

This tool requires the memory coordinator to be initialized (Redis must be connected). The confidence score ranges from 0.0 to 1.0. The sources array lists episodic memory IDs used to generate the answer.

Discovery Tools

discover_tools — Search Tool Registry

Search the tool registry for available tools by keyword or category. This meta-tool is always included in the LLM context window (regardless of dynamic tool selection) so the agent can discover tools that were not pre-selected for the current request.

Input Schema

query

string

required

Keyword search query to match against tool names and descriptions.

Example Input

{
  "query": "temperature",
  "category": "smart_home"
}

Success Response

[
  {
    "name": "home.set_temperature",
    "description": "Set the thermostat temperature for a room",
    "category": "smart_home"
  },
  {
    "name": "home.get_temperature",
    "description": "Read the current temperature from a room sensor",
    "category": "smart_home"
  }
]

Error Response

Error: Tool selector not configured

This tool is only registered when a toolSelector is configured (dynamic tool selection is enabled). It allows the agent to search for tools that were filtered out by the initial similarity-based selection. See Dynamic Tool Selection for details.

Connecting to the MCP Server

Using the MCP SDK (Node.js)

import { Client } from '@modelcontextprotocol/sdk/client/index.js';
import { StreamableHTTPClientTransport } from '@modelcontextprotocol/sdk/client/streamableHttp.js';

const transport = new StreamableHTTPClientTransport(
  new URL('http://localhost:4100/mcp')
);

const client = new Client({
  name: 'my-client',
  version: '1.0.0',
});

await client.connect(transport);

// Call a tool
const result = await client.callTool({
  name: 'system.health',
  arguments: {},
});

console.log(result.content);

Using curl (for testing)

The MCP protocol uses JSON-RPC over HTTP. You can test with curl by sending properly formatted JSON-RPC requests:

# Initialize the MCP session
curl -X POST http://localhost:4100/mcp \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "initialize",
    "params": {
      "protocolVersion": "2025-03-26",
      "capabilities": {},
      "clientInfo": { "name": "curl-test", "version": "1.0.0" }
    }
  }'

Claude Desktop Configuration

Add agtOS as an MCP server in your Claude Desktop config (claude_desktop_config.json):

{
  "mcpServers": {
    "agtos": {
      "url": "http://localhost:4100/mcp"
    }
  }
}

Tool Availability

Not all tools are always available. Tool registration depends on which services are running:

Tool Group	Dependency	Condition
Voice tools	Voice pipeline	Always registered (may return errors if pipeline down).
System tools	Health manager	Always registered.
Memory tools	Memory coordinator	Registered only if memory coordinator exists (Redis connected).
Workflow tools	Workflow engine	Registered only if workflow engine exists.
Scheduler tools	Redis scheduler	Registered only if Redis is connected.
Discovery tools	Tool selector	Registered only if dynamic tool selection is configured.

Use tools/list to discover which tools are currently available on a running server.

​Voice Tools

​Input Schema

​Example Input

​Success Response

​Error Response

​Input Schema

​Example Input

​Success Response

​Error Response

​System Tools

​Input Schema

​Success Response

​Error Response

​Input Schema

​Example Input (all sessions)

​Example Input (specific session)

​Success Response (all sessions)

​Error Response

​Orchestration Tools

​Input Schema

​Example Input

​Success Response

​Error Response

​Input Schema

​Success Response

​Error Response

​Input Schema

​Example Input (cron)

​Example Input (once)

​Example Input (interval)

​Success Response

​Error Response

​Input Schema

​Example Input

​Success Response

​Error Response

​Input Schema

​Example Input

​Success Response

​Error Response

​Memory Tools

​Input Schema

​Example Input

​Success Response

​Error Response

​Discovery Tools

​Input Schema

​Example Input

​Success Response

​Error Response

​Connecting to the MCP Server

​Using the MCP SDK (Node.js)

​Using curl (for testing)

​Claude Desktop Configuration

​Tool Availability

Voice Tools

Input Schema

Example Input

Success Response

Error Response

Input Schema

Example Input

Success Response

Error Response

System Tools

Input Schema

Success Response

Error Response

Input Schema

Example Input (all sessions)

Example Input (specific session)

Success Response (all sessions)

Error Response

Orchestration Tools

Input Schema

Example Input

Success Response

Error Response

Input Schema

Success Response

Error Response

Input Schema

Example Input (cron)

Example Input (once)

Example Input (interval)

Success Response

Error Response

Input Schema

Example Input

Success Response

Error Response

Input Schema

Example Input

Success Response

Error Response

Memory Tools

Input Schema

Example Input

Success Response

Error Response

Discovery Tools

Input Schema

Example Input

Success Response

Error Response

Connecting to the MCP Server

Using the MCP SDK (Node.js)

Using curl (for testing)

Claude Desktop Configuration

Tool Availability