AI Context Exchange Protocol
Agents hand off work across tools and roles. AICEP’s rules keep that handoff clean: secrets, private data, and agents’ jobs stay where they belong.
Agent harness
Grok Bot
Pass context in and out of Grok Bot.
Agent harness
MetaDots
Pass context in and out of MetaDots, Meta AI’s agent harness.
Agent harness
Claude Code / Claude Desktop
Pass context in and out of Claude Code and Claude Desktop.
Agent harness
OpenAI Codex / ChatGPT agent
Pass context in and out of OpenAI Codex and ChatGPT agent.
Agent harness
Cursor
Pass context in and out of Cursor’s coding agents.
Agent harness
OpenClaw
Pass context in and out of OpenClaw.
Agent harness
Your own MCP or agent harness
Bring your own harness: any MCP server or custom agent runtime can seat AICEP.
Seat AICEP in your harness →AICEP 0.1 — Public Draft (RFC)
The AI Context Exchange Protocol is published for comment. Classify before you send, seat an agent before it talks, keep credentials off the bus.
Contribute
We’d love to hear from anyone interested in helping develop this protocol or contributing to it. Comments, implementation notes and questions are welcome.
Latest updates
- In progressNext revision: taking AICEP from individual use to corporate use, incorporating more elements of corporate cybersecurity standards into the protocol.
- Companion runbook v2.5.1: each sending agent now publishes into its own folder, and a receiver reads only the folders of senders it has verified and not revoked; a batch signed by a different sender than the one registered for its folder is rejected and set aside for review. A batch's validity window is capped at seven days, the receiver's audit log refuses to grow if its tail has been cut, and only a named person can re-anchor it or approve a gap. Follow-up fixes cap how many unknown folders the receiver logs, combine repeated folder-mismatch alerts, and take a held batch's folder from the registry. 188 automated tests, all passing.
- Companion runbook v2.4.1: the proposed Phase 2 encryption extension adds a signed verifier attestation; 90 automated tests, all passing; new G11 prompt (step 2.8).
- AICEP 0.1 published as a Public Draft (RFC), open for comment.
Get started
- Read the Public Draft (RFC).
- Point your coordinator agent at the import / seat page to seat AICEP.
- Work through the companion runbook: prompts, paper tests, the encryption extension and reference code.
- See use cases (library coming).
- Send comments to hello@aicep.ai.
About AICEP
AICEP shares typed context with fences in the protocol itself. Classify before you send. Seat an agent before it talks. Keep credentials off the bus. Exchange second.
Not a chatbot standard. Not a certification. Published by AIOptiTech.
AICEP was created by AIOptiTech.
Built for agent harnesses
AICEP is harness-neutral: it is meant to pass context in and out of the agent harnesses people already run.
Grok Bot
Pass context in and out of Grok Bot.
MetaDots
Pass context in and out of MetaDots, Meta AI’s agent harness.
Claude Code / Claude Desktop
Pass context in and out of Claude Code and Claude Desktop.
OpenAI Codex / ChatGPT agent
Pass context in and out of OpenAI Codex and ChatGPT agent.
Cursor
Pass context in and out of Cursor’s coding agents.
OpenClaw
Pass context in and out of OpenClaw.
Your own MCP or agent harness
Bring your own harness: any MCP server or custom agent runtime can seat AICEP.
Seat AICEP in your harness →