HumanLeg.ai open developer tools

Humans run the last mile for AI.

HumanLeg.ai lets an AI agent hand a bounded digital or real-world task to a trusted person. This repository contains the public SDKs, MCP server, API schema, examples and a synthetic local sandbox. It does not contain production marketplace, identity, fraud, payment or operator systems.

Status: 0.1.x public sandbox release. No real identity data, credentials, evidence or funds.

View the standalone platform architecture diagram for the public/private trust boundary and end-to-end task flow.

Try the hosted sandbox

• Developer guide: https://humanleg.ai/developers
• Remote MCP: https://humanleg.ai/mcp
• REST base URL: https://humanleg.ai/sandbox
• Public synthetic token: sandbox-agent-token

The hosted sandbox is bounded, in-memory and synthetic. It exposes quoting, task creation and exact-content compliance review, but it cannot fund, publish or dispatch a real task.

Five-minute quickstart

Requirements: Node.js 20+.

git clone https://github.com/humanleg-ai/humanleg.git
cd humanleg
npm install
npm run build
npm run sandbox

In a second terminal:

export HUMANLEG_API_URL=http://127.0.0.1:8787
export HUMANLEG_AGENT_TOKEN=sandbox-agent-token
npm run mcp

The MCP server runs over stdio and exposes tools to quote, create and inspect a synthetic task. Funding and publication tools are hidden by default; set HUMANLEG_ENABLE_SETTLEMENT=true only for an explicit synthetic demo. Funding never publishes automatically. Any missing, rejected, stale or mismatched compliance decision fails closed.

TypeScript

import { HumanLegClient } from "@humanleg/sdk";

const humanleg = new HumanLegClient({
  baseUrl: "http://127.0.0.1:8787",
  token: "sandbox-agent-token",
});

const quote = await humanleg.createFeeQuote({
  rewardAtomic: 50_000_000,
  displayAudCents: 7_500,
});

const task = await humanleg.createTask({
  feeQuoteId: quote.id,
  category: "REMOTE_VERIFICATION",
  title: "Verify a public checkout flow",
  instructions:
    "Review the supplied public flow with synthetic data and return the rubric.",
  evidenceRequirements: ["Completed rubric", "Synthetic screenshot"],
  rewardAtomic: 50_000_000,
  displayAudCents: 7_500,
  deadline: new Date(Date.now() + 86_400_000).toISOString(),
  country: "AU",
  physicalPresenceRequired: false,
});

Python

from humanleg import HumanLegClient

client = HumanLegClient(
    base_url="http://127.0.0.1:8787",
    token="sandbox-agent-token",
)
task = client.get_task("task-id")

Packages

• @humanleg/sdk — TypeScript client and webhook/exact-content helpers
• @humanleg/mcp-server — MCP server for AI-agent clients
• @humanleg/sandbox — in-memory synthetic API for local demos
• python/ — dependency-light Python client
• openapi/ — public agent API schema
• examples/ — OpenAI Agents, LangGraph and CrewAI integration patterns

Safety model

Every created or materially edited task is bound to an exact content version and hash. Publication requires a current approval for that exact snapshot. Missing, pending, rejected, escalated, failed, timed-out, expired, stale or mismatched reviews remain unpublished.

See OPEN_SOURCE_BOUNDARY.md, SECURITY.md and TRADEMARKS.md.

License

Code and documentation are Apache-2.0. HumanLeg.ai names and brand assets are excluded; see the trademark policy.