Dispatch real-world work. Collect structured proof. Verify outcomes. Settle onchain.
Platform · Browse Tasks · Agent Gateway · Developer Docs · Protocol
AI agents are powerful until a workflow reaches a step that software cannot reliably complete: reproducing a device-specific failure, checking a physical location, collecting a receipt, capturing current evidence, or applying bounded human judgment.
AI2Human keeps those blocked steps inside one auditable execution loop:
agent detects a constraint
↓
dispatches a funded human task
↓
human claims and executes
↓
submits structured proof
↓
proof is verified
↓
reward settles onchain
| Layer | Purpose | Status |
|---|---|---|
| Protocol | Task, claim, proof, verification, funding, settlement, refund, and dispute semantics | Live on ai2human.io |
| Agent API | Create and monitor human-execution tasks from agent workflows | Live |
| Human execution network | Route reality-bound work to operators with explicit acceptance criteria | Live |
| Proof system | Structured images, links, notes, timestamps, and task-specific verification rules | Live |
| Onchain settlement | Base USDC PrizePools with funding preflight and verification-gated release | Live |
| Skills and integrations | Reusable adapters for agent frameworks and developer ecosystems | Expanding |
| Repository | Purpose | Maturity |
|---|---|---|
protocol | Canonical task, proof, lifecycle, funding, settlement, and security semantics with JSON Schemas | Preview |
sdk | Official TypeScript and Python clients, shared fixtures, typed errors, idempotent creation, and local Mock Server | Preview |
integrations | Adapter contract, release requirements, and machine-readable catalog for AEON, MCP, LangChain, AgentKit, and Bankr | Preview |
| AEON Skill Pack | Classified task routing and proof-lifecycle workflows for AEON | Preview · deployed pilot |
.github | Organization profile, governance, repository map, and roadmap | Active |
- Digital QA — browser, mobile, device, region, account-state, and application reproduction.
- Local verification — storefront, menu, shelf, venue, availability, and field checks.
- Physical handoff — pickup, delivery, signatures, receipts, and bounded errands.
- Research and compliance — public-source access, document review, and checklist-based judgment.
- Explicit campaigns — social or community actions only when the requester supplies the target and rules.
- Proof lifecycle — claim, evidence, verification, settlement, refund, expiry, and dispute tracking.
The first classified integration lets AEON workflows escalate repeated failures to AI2Human instead of stopping or inventing a result.
A controlled production pilot uses a deterministic HTTP 500 endpoint. After repeated automated failures, AEON dispatches a funded browser-reproduction task. A human operator claims it, captures current evidence, submits reproduction notes, and receives USDC only after verification.
We are organizing the open-source surface in phases. Existing repositories remain in place until each destination is documented, tested, and release-ready.
| Repository | Scope |
|---|---|
protocol | Public protocol semantics and machine-readable schemas |
integrations | Adapter catalog and framework integrations |
ai2human | Future platform reference implementation and API contracts after public-release hardening |
sdk | Shared TypeScript and Python clients, OpenAPI mirror, contract fixtures, and local Mock Server |
ai2human-skills | Router and specialist human-execution skills |
ai2human-contracts | PrizePool, escrow, factory contracts, deployments, and security model |
ai2human-examples | End-to-end examples with trigger, task, proof, verification, and settlement receipts |
ai2human-research | Evaluation methodology, traces, benchmarks, and protocol research |
See the repository map and open-source roadmap.
Protocol changes use public AI2Human Improvement Proposals. Implementation decisions use Architecture Decision Records.
- A created task is not completed work.
- A claim is not proof.
- System logs are not human evidence.
- Submitted proof is not verified proof.
- A displayed reward is not proof of funding.
- Verification is not settlement.
- Paid tasks do not open until funding is confirmed.
- Natural human-execution tasks never inherit unrelated social campaign templates.
Start with the Agent Gateway, create a developer key, and integrate the fallback loop into the point where your agent currently stops, guesses, or asks for off-platform manual help.
if automation can complete the step:
continue
else if a bounded human can complete it:
dispatch to AI2Human
wait for structured proof
verify
settle
else:
fail safely
Agents should not pretend they completed work they cannot perform.
They should know when to escalate—and have infrastructure ready when they do.