[Idea]: Any thought on using Agent Client Protocol?

[kasajian]

What are you trying to do?

Support:
Using an existing subscription via:
https://agentclientprotocol.com/get-started/introduction

Use a library such as:
https://agentclientprotocol.github.io/python-sdk/

What feels difficult or missing today?

ability to use an existing subscription

What outcome would help?

support for ACP

How do you handle this today?

Unable to use the product without it

Additional context, examples, or possible directions

Many enterprises have a single subscription... such as Codex, Claude Code, Copilot, etc. Users can only use those tools. They don't have API keys they can plug in. However, those subscriptions can be used through tools that use the Agent Client Protocol to access LLM rather than the LLM OpenAI / Anthropic compatible endpoints directly. This is not about cost but legality and indemnification (https://blogs.microsoft.com/on-the-issues/2023/09/07/copilot-copyright-commitment-ai-legal-concerns/) of results produced by AI.

How would you like to participate?

• I can add use cases or join the product/design discussion.
• I can help explore the technical approach.
• I can help implement an approved work item.
• I can test or validate a proposed solution.

Before posting

• I searched existing Discussions and Issues for related proposals.

[Sure-Will]

@kasajian +1 from a self-hosted Open Notebook user: I already pay for Codex/ChatGPT Pro, and maintaining a separate API key and usage balance makes Open Notebook much more expensive to use. ACP looks like a useful common layer; for Codex, its official App Server/SDK already supports managed ChatGPT login, so I would be happy to test a local, single-user integration that respects subscription limits, never exports or shares tokens, stays off the public network, and keeps embeddings as a separate provider when needed.

[lfnovo]

ACP isn't the right layer for this: it's an editor↔coding-agent protocol (turns, tool calls, file-system callbacks), not an inference API — no completions, structured output or embeddings, which is what our model layer needs. And the subscription part is blocked by the vendors' terms, not by us.

The route we're working on solves your actual problem better, and the other way around: an MCP server for Open Notebook (#878). You keep using Claude or Codex with your own subscription, exactly as your company allows, and it calls Open Notebook behind the scenes for search, sources and evidence. Reasoning and its legal terms stay with the vendor product you're already licensed for; Open Notebook only does retrieval. Server-side jobs that need a model (source processing, embeddings) still need a configured provider — a local one like Ollama works without any API key.

@Sure-Will, that's where your testing offer would land once #878 has something to run.

Status: answer — ACP not planned; the need is routed to the MCP server (#878). Closing this thread; #878 is where to follow.