Configuration reference

Every option createAgent() accepts, in one place. For what each type is for, see agent types.

Shared options#

Accepted by all three types.

Option	Type	Default	Notes
`type`	`'one-shot' \| 'agent' \| 'saga'`	required	Selects the overload
`systemPrompt`	`string`	-	The agent's role and quality bar
`model`	`string`	type-dependent	See the defaults below
`effort`	`string`	model default	Reasoning effort, e.g. `xhigh`
`maxOutputTokens`	`number`	`128000`	Cap on generated tokens

By type#

Option	`one-shot`	`agent`	`saga`
`temperature`	✓	-	-
`stop`	✓	-	-
`instructions`	-	✓	✓
`deadlineMs`	-	✓	✓
`contextWindow`	-	✓	✓
`tools`	-	✓	✓
`toolAllowlist`	-	✓	✓
`maxTools`	-	✓	✓
`skills`	-	✓	✓
`mcpServers`	-	✓	✓
`imageModel`	-	✓	✓
`outputs`	-	✓	✓
`maxOutputInlineBytes`	-	✓	✓
`maxTurns`	-	✓	-
`permissionMode`	-	✓	-
`groundingModel`	-	-	✓
`maxSteps`	-	-	✓
`parallelism`	-	-	✓

TypeScript enforces this table. An option in the wrong column is a compile error, not a runtime surprise.

Defaults#

Option	Default	What it sets
`model` for `one-shot`	`deepseek/deepseek-v4-flash`	Model used by the synchronous response endpoint
`model` for `agent` and `saga`	`deepseek/deepseek-v4-pro`	Model used by the managed runtime
`maxOutputTokens`	`128000`	Maximum tokens the model may generate in one response
`contextWindow`	`250000`	Size of the agent's working context across a run
`deadlineMs`	none	Wall-clock execution deadline. Unset means the run is unbounded
`permissionMode`	`auto`	Default permission policy for `agent`
`groundingModel`	`deepseek/deepseek-v4-flash`	Grounding model for `saga`
`maxOutputInlineBytes`	`262144`	Maximum bytes projected inline when not set explicitly
`toolAllowlist`	derived	Tool names visible to the model. Derived from the rest of the config when unset. See Tools

Connection options#

The second argument configures transport rather than behaviour, and is the same shape new CoreSource() takes.

Option	Type	Default
`apiKey`	`string`	`CORESOURCE_API_KEY`
`baseUrl`	`string \| URL`	`CORESOURCE_API_URL`, then `https://api.coresource.ai/v1/`
`fetch`	`typeof fetch`	global `fetch`
`ttlSeconds`	`number`	`3600` for an SDK-hosted bridge

ttlSeconds applies only when local capabilities cause the SDK to provision a bridge. The accepted range is 60 seconds through 30 days.

TypeScript

import { createAgent } from '@coresourceai/sdk';

const agent = createAgent(
  { type: 'agent', systemPrompt: 'Be careful.' },
  {
    apiKey: process.env.CORESOURCE_API_KEY,
    baseUrl: 'https://api.example.com/v1/'
  }
);

await agent.close();

See Authentication for keys, deployments, and supplying your own fetch.

Lifecycle#

createAgent() is synchronous and lazy. It prepares the agent locally and does no network work until run() or start().

close() releases what the agent holds and makes further use of the handle an error. A live bridge is created when an agent or saga uses custom tools, runtime tools, local skill bundles, or locally registered MCP.

TypeScript

import { createAgent } from '@coresourceai/sdk';

const agent = createAgent({ type: 'agent' });

try {
  const result = await agent.run('Do the work.');
  console.log(result.text);
} finally {
  await agent.close();
}

registerRuntimeTools() and registerMcpServer() must be called before the first run starts; afterwards they throw, because the tool set is fixed when the run begins.