The Ainfera-native agent framework. MCP-first. A2A-first. Trust-scored by default.
Status: Design · Alpha arriving Q3 2026
OpenClaw is the Ainfera-native agent framework. It is optimized for:
- MCP-first — tools defined as Model Context Protocol servers by default, not as framework-specific adapters.
- A2A-first — agents as peers with verifiable W3C DID identities, not monoliths with embedded tools.
- Trust-scored by default — every invocation is ATS-scored, sandboxed, and metered via the Ainfera SDK. No opt-in, no configuration.
- Claude-first ergonomics — MCP and the Anthropic SDK are the primary adapters. Other model providers via adapter layer in Q4 2026.
This is the design repository. It contains:
- Early design spec (docs/spec-v0.md)
- Roadmap and principles
- Public comment surface (open an Issue to discuss)
OpenClaw is not yet released. The framework will be available as
pip install openclaw in Q3 2026.
In the meantime, Ainfera's trust scoring, sandboxing, and metered billing work with every major agent framework via the Ainfera SDK:
ainfera— CLI (pip install ainfera)ainfera-sdk— Python SDK (pip install ainfera-sdk)
Existing frameworks (LangChain, CrewAI, AutoGen, LlamaIndex) optimize for rapid prototyping of agentic workflows. They are excellent for that. OpenClaw optimizes for three things those frameworks deprioritize:
| OpenClaw priority | What others do |
|---|---|
| MCP as the default tool protocol | Framework-specific tool decorators |
A2A agent cards at /.well-known/agent-card.json |
Agents as libraries, not peers |
| ATS trust score embedded per invocation | Trust as optional instrumentation |
| Sandboxing + kill switch by default | Execution isolation as afterthought |
Every agent gets:
- W3C DID identity (
did:ainfera:agent:{uuid}) on registration - MCP server exposing its tools at
/mcp/{agent-name} - A2A agent card at
/.well-known/agent-card.json - Automatic ATS trust scoring from execution telemetry
- Redis-flagged kill switch, checked every 100ms (sub-500ms SLA)
No environment variables. No configuration files. Working defaults.
Tools are defined as MCP servers, not as framework-specific classes. A tool authored for one OpenClaw agent runs in Claude Desktop, Cursor, the Ainfera marketplace, or any MCP client — unchanged.
Agent-to-agent communication happens over the Agent2Agent (A2A) protocol, not over framework-specific calling conventions. Agents can discover, authenticate, and invoke each other regardless of framework.
The primary adapter layer targets Anthropic's SDK and MCP. Adapter layers for OpenAI, Google, NIM, and open-weight models (via Hugging Face) arrive in Q4 2026.
| Quarter | Milestone |
|---|---|
| Q2 2026 | Spec v0 (this repo) · community comment window |
| Q3 2026 | v0.1 alpha · single-agent use cases · Ainfera SDK integration |
| Q4 2026 | v0.2 alpha · A2A multi-agent · Cedar + OPA policy engine · OpenAI/NIM adapters |
| Q1 2027 | v0.5 beta · LangChain/CrewAI migration shim package |
| Q2 2027 | v1.0 stable |
OpenClaw explicitly does not aim to be:
- A general-purpose agent framework (use LangChain/CrewAI)
- A model provider abstraction layer (use LiteLLM/OpenRouter)
- A workflow orchestrator (use Temporal/Prefect)
- An observability platform (use LangSmith/W&B)
OpenClaw is narrow on purpose: MCP tools, A2A peers, ATS-scored execution. That's the whole thing.
The design is open for comment. Please:
- Read docs/spec-v0.md
- Open an Issue with any of:
- Critique of design decisions
- Use cases not covered
- Protocol conflicts with A2A or MCP
- Naming suggestions (the name is not final)
Pull requests during the design phase are accepted only for spec clarifications and typos. Implementation PRs open in Q3 2026.
- Model Context Protocol (MCP) — Anthropic
- Agent2Agent (A2A) — Google (Linux Foundation stewardship)
- W3C Decentralized Identifiers (DIDs)
- Verifiable Credentials Data Model
Apache 2.0 — see LICENSE.
Hizrian Raz — founder, Ainfera Pte. Ltd.
© 2026 Ainfera Pte. Ltd. · Singapore