Ephemeral agents

An ephemeral agent is described by the request rather than persisted as a row in MANTYX’s Agent table. The full spec (system prompt, model, tools) is stored as part of each session/run for observability but is not editable from the dashboard.

This is the right choice when:

• You’re prototyping or scripting and don’t need a stable agent across runs.
• You want maximum control over the system prompt at the call site.
• The set of tools changes per call.

Minimal one-shot

const result = await client.runAgent({
  systemPrompt: "You are a helpful assistant.",
  prompt: "What's the capital of France?",
});
console.log(result.text);

Adding tools

Seven flavours, all carried inside the agent spec. The split is along who can reach the resource — server-resolved tools run inside MANTYX; client-resolved tools run inside your SDK process and shuttle results back over the agent loop.

kind	Resolved by	Notes
mantyx	server	A workspace Tool row referenced by id. See MANTYX tools.
mantyx_plugin	server	A platform plugin tool referenced by @plugin/tool name. See Plugin tools.
a2a	server	A remote Agent2Agent peer MANTYX dials directly.
mcp	server	A remote MCP server (Streamable HTTP) MANTYX lists and proxies.
local	client	Defined and executed inside your SDK process. See Local tools.
a2a_local	client	An A2A peer only the SDK can reach (intranet, on-device).
mcp_local	client	An MCP server only the SDK can reach (stdio, intranet).

import { defineLocalTool, mantyxTool, mantyxPluginTool } from "@mantyx/sdk";
import { z } from "zod";

await client.runAgent({
  systemPrompt: "You are a research assistant.",
  prompt: "Look up the latest CPI release and summarise it.",
  tools: [
    mantyxPluginTool("@web/search"),
    mantyxTool("tool_cm6abc123"),
    defineLocalTool({
      name: "save_note",
      parameters: z.object({ title: z.string(), body: z.string() }),
      execute: async ({ title, body }) => {
        // ...write to disk
        return "ok";
      },
    }),
  ],
});

Picking a model

Pass modelId (TypeScript / Python) or ModelID (Go) to override the workspace default. See Models for the supported shorthand syntax.

Tuning thinking effort

Pass reasoningLevel (TypeScript) / reasoning_level (Python) / ReasoningLevel (Go) to dial provider extended thinking on reasoning models. The value is forwarded unchanged to the server, which maps it onto each LLM’s native dial. Accepts a string anchor ("off", "low", "medium", "high") or an integer in [0, 100] — see Reasoning level for the full table.

await client.runAgent({
  systemPrompt: "...",
  prompt: "Plan a multi-week migration.",
  reasoningLevel: "high",
});

Budgeting tool turns

await client.runAgent({
  systemPrompt: "...",
  prompt: "...",
  budgets: { maxToolTurns: 8 }, // hard cap
});

If the model wants to call tools more than maxToolTurns times, the run terminates with result.subtype = "error_max_tool_turns".