OpenServ Agent SDK

OpenServ Agent SDK

Build and deploy custom AI agents for the OpenServ platform using TypeScript.

Why build an agent?

An OpenServ agent is a service that runs your code and exposes it on the OpenServ platform—so it can be triggered by workflows, other agents, or paid calls (e.g. x402). The platform sends tasks to your agent; your agent runs your capabilities (APIs, tools, file handling) and returns results. You don't have to use an LLM—e.g. it could be a static API that just returns data. If you need LLM reasoning, you have two options: use runless capabilities (the platform handles the AI call for you—no API key needed) or use generate() (delegates the LLM call to the platform); alternatively, bring your own LLM (any provider you have access to).

How it works (the flow)

1. Define your agent — System prompt plus capabilities. Capabilities come in two flavors: runnable (with a Zod schema and a run handler) and runless (just a name and description—the platform handles the AI call automatically). You can also use generate() inside runnable capabilities to delegate LLM calls to the platform.
2. Register with the platform — You need an account on the platform; often the easiest way is to let provision() create one for you automatically by creating a wallet and signing up with it (that account is reused on later runs). Call provision() (from @openserv-labs/client): it creates or reuses a wallet, registers the agent, and writes API key and auth token into your env (or you pass agent.instance to bind them directly). In development you can skip setting an endpoint URL; the SDK can use a built-in tunnel to the platform.
3. Start the agent — Call run(agent). The agent listens for tasks, runs your capabilities (and your LLM if you use one), and responds. Use reference.md and troubleshooting.md for details; examples/ has full runnable code.

What your agent can do

• Runless Capabilities — Just a name and description. The platform handles the AI call automatically—no API key, no run() function needed. Optionally define inputSchema and outputSchema for structured I/O.
• Runnable Capabilities — The tools your agent can run (e.g. search, transform data, call APIs). Each has a name, description, inputSchema, and run() function.
• generate() method — Delegate LLM calls to the platform from inside any runnable capability. No API key needed—the platform performs the call and records usage. Supports text and structured output.
• Task context — When running in a task, the agent can attach logs and uploads to that task via methods like addLogToTask() and uploadFile().
• Multi-agent workflows — Your agent can be part of workflows with other agents; see the openserv-client skill for the Platform API, workflows, and ERC-8004 on-chain identity.

Reference: reference.md (patterns) · troubleshooting.md (common issues) · examples/ (full examples)

Quick Start

Installation

npm install @openserv-labs/sdk @openserv-labs/client zod

Note: openai is only needed if you use the process() method for direct OpenAI calls. Most agents don't need it—use runless capabilities or generate() instead.

Minimal Agent

See examples/basic-agent.ts for a complete runnable example.

The pattern is simple:

1. Create an Agent with a system prompt
2. Add capabilities with agent.addCapability()
3. Call provision() to register on the platform (pass agent.instance to bind credentials)
4. Call run(agent) to start

Complete Agent Template

File Structure

my-agent/
├── src/agent.ts
├── .env
├── .gitignore
├── package.json
└── tsconfig.json

Dependencies

npm init -y && npm pkg set type=module
npm i @openserv-labs/sdk @openserv-labs/client dotenv zod
npm i -D @types/node tsx typescript

Note: The project must use "type": "module" in package.json. Add a "dev": "tsx src/agent.ts" script for local development. Only install openai if you use the process() method for direct OpenAI calls.

.env

Most agents don't need any LLM API key—use runless capabilities or generate() and the platform handles LLM calls for you. If you use process() for direct OpenAI calls, set OPENAI_API_KEY. The rest is filled by provision().

# Only needed if you use process() for direct OpenAI calls:
# OPENAI_API_KEY=your-openai-key
# ANTHROPIC_API_KEY=your_anthropic_key  # If using Claude directly

# Required for deploy (get from OpenServ platform dashboard)
OPENSERV_USER_API_KEY=your-user-api-key

# Auto-populated by provision():
WALLET_PRIVATE_KEY=
OPENSERV_API_KEY=
OPENSERV_AUTH_TOKEN=
PORT=7378
# Production: skip tunnel and run HTTP server only
# DISABLE_TUNNEL=true
# Force tunnel even when endpointUrl is set
# FORCE_TUNNEL=true

Capabilities

Capabilities come in two flavors:

Runless Capabilities (recommended for most use cases)

Runless capabilities don't need a run function—the platform handles the AI call automatically. Just provide a name and description:

// Simplest form — just name + description
agent.addCapability({
  name: 'generate_haiku',
  description: 'Generate a haiku poem (5-7-5 syllables) about the given input.'
})

// With custom input schema
agent.addCapability({
  name: 'translate',
  description: 'Translate text to the target language.',
  inputSchema: z.object({
    text: z.string(),
    targetLanguage: z.string()
  })
})

// With structured output
agent.addCapability({
  name: 'analyze_sentiment',
  description: 'Analyze the sentiment of the given text.',
  outputSchema: z.object({
    sentiment: z.enum(['positive', 'negative', 'neutral']),
    confidence: z.number().min(0).max(1)
  })
})

• No run function — the platform performs the LLM call
• No API key needed — the platform handles it
• inputSchema is optional — defaults to z.object({ input: z.string() }) if omitted
• outputSchema is optional — define it for structured output from the platform

See examples/haiku-poet-agent.ts for a complete runless example.

Runnable Capabilities

Runnable capabilities have a run function for custom logic. Each requires:

• name - Unique identifier
• description - What it does (helps AI decide when to use it)
• inputSchema - Zod schema defining parameters
• run - Function returning a string

agent.addCapability({
  name: 'greet',
  description: 'Greet a user by name',
  inputSchema: z.object({ name: z.string() }),
  async run({ args }) {
    return `Hello, ${args.name}!`
  }
})

See examples/capability-example.ts for basic capabilities.

Note: The schema property still works as an alias for inputSchema but is deprecated. Use inputSchema for new code.

Using Agent Methods

Access this in capabilities to use agent methods like addLogToTask(), uploadFile(), generate(), etc.

See examples/capability-with-agent-methods.ts for logging and file upload patterns.

Agent Methods

generate() — Platform-Delegated LLM Calls

The generate() method lets you make LLM calls without any API key. The platform performs the call and records usage to the workspace.

// Text generation
const poem = await this.generate({
  prompt: `Write a short poem about ${args.topic}`,
  action
})

// Structured output (returns validated object matching the schema)
const metadata = await this.generate({
  prompt: `Suggest a title and 3 tags for: ${poem}`,
  outputSchema: z.object({
    title: z.string(),
    tags: z.array(z.string()).length(3)
  }),
  action
})

// With conversation history
const followUp = await this.generate({
  prompt: 'Suggest a related topic.',
  messages,  // conversation history from run function
  action
})

Parameters:

• prompt (string) — The prompt for the LLM
• action (ActionSchema) — The action context (passed into your run function)
• outputSchema (Zod schema, optional) — When provided, returns a validated structured output
• messages (array, optional) — Conversation history for multi-turn generation

The action parameter is required because it identifies the workspace/task for billing. Use it inside runnable capabilities where action is available from the run function arguments.

Task Management

await agent.createTask({ workspaceId, assignee, description, body, input, dependencies })
await agent.updateTaskStatus({ workspaceId, taskId, status: 'in-progress' })
await agent.addLogToTask({ workspaceId, taskId, severity: 'info', type: 'text', body: '...' })
await agent.markTaskAsErrored({ workspaceId, taskId, error: 'Something went wrong' })
const task = await agent.getTaskDetail({ workspaceId, taskId })
const tasks = await agent.getTasks({ workspaceId })

File Operations

const files = await agent.getFiles({ workspaceId })
await agent.uploadFile({ workspaceId, path: 'output.txt', file: 'content', taskIds: [taskId] })
await agent.deleteFile({ workspaceId, fileId })

Action Context

The action parameter in capabilities is a union type — task only exists on the 'do-task' variant. Always narrow with a type guard before accessing action.task:

async run({ args, action }) {
  // action.task does NOT exist on all action types — you must narrow first
  if (action?.type === 'do-task' && action.task) {
    const { workspace, task } = action
    workspace.id        // Workspace ID
    workspace.goal      // Workspace goal
    task.id             // Task ID
    task.description    // Task description
    task.input          // Task input
    action.me.id        // Current agent ID
  }
}

Do not extract action?.task?.id before the type guard — TypeScript will error with Property 'task' does not exist on type 'ActionSchema'.

Workflow Name & Goal

The workflow object in provision() requires two important properties:

• name (string) - This becomes the agent name in ERC-8004. Make it polished, punchy, and memorable — this is the public-facing brand name users see. Think product launch, not variable name. Examples: 'Crypto Alpha Scanner', 'AI Video Studio', 'Instant Blog Machine'.
• goal (string, required) - A detailed description of what the workflow accomplishes. Must be descriptive and thorough — short or vague goals will cause API calls to fail. Write at least a full sentence explaining the workflow's purpose.

workflow: {
  name: 'Haiku Poetry Generator',  // Polished display name — the ERC-8004 agent name users see
  goal: 'Transform any theme or emotion into a beautiful traditional 5-7-5 haiku poem using AI',
  trigger: triggers.x402({ ... }),
  task: { description: 'Generate a haiku about the given topic' }
}

Trigger Types

import { triggers } from '@openserv-labs/client'

triggers.webhook({ waitForCompletion: true, timeout: 600 })
triggers.x402({ name: '...', description: '...', price: '0.01', timeout: 600 })
triggers.cron({ schedule: '0 9 * * *' })
triggers.manual()

Important: Always set timeout to at least 600 seconds (10 minutes) for webhook and x402 triggers. Agents often take significant time to process requests — especially when performing research, content generation, or other complex tasks. A low timeout will cause premature failures. For multi-agent pipelines with many sequential steps, consider 900 seconds or more.