This protocol is implemented by zed industries (https://agentclientprotocol.com), for communication between editors and genai agents.
It’s implemented in zed obviously (https://zed.dev/blog/bring-your-own-agent-to-zed) and in neovim, so I wanted to suggest it as an idea for sublime text also, in a manner as LSP-copilot chat, for example.
It looks like there is significant overlap with the Model Context Protocol (MCP) which is used by VS Code (I might be wrong because I didn’t take a very close look yet).
I think the biggest obstacle for a Sublime plugin is that the UI is restricted by what is available in the ST API (input handlers, output panels, HTML sheets, …) and therefore they have to fiddle with all kinds of workarounds to utilize these generic components for the UI. On the other hand VS Code and Zed seem to intend to make the AI stuff directly a part of the editor, so they can put a lot of effort in implementing UI components that are specifically designed for the communication with AI agents / LLM.
I don’t know about ACP but MCP is a security nightmare. Nobody should be using that on their computer in its current state. It’s horrifying that anyone does. I’ll assume after a brief search that ACP creates similar vulnerabilities.
It’s not my business if someone wants to punch a Boeing-sized security hole in their Sublime Text editor by installing a dangerous plugin, but I certainly hope Sublime HQ does not spend any developer cycles supporting such endeavors in the core editor, including making such plugins easier to create. There are way too many other more productive things they could be doing.
Yes, it is similar to Anthropic’s MCP, it’s the same idea with little overhead to make it work.
Besides Gemini CLI and Claude code, Qwen code and opencode are already in implementation phase
A minimal interface such as is provided by the aforementioned LSP-copilot would suffice for most user cases, and AFAIK it can be done without such ACP implementation.
Particularly, a more generic version that is able to use local models would be a great extension, now that
Yollama seems to have died.
I wanted to share my concrete implementation that I’ve been working on: GeminiCLI . It integrates the google/gemini-cli directly into Sublime Text using its --acp flag. It demonstrates how we can effectively leverage agent cli in sublime text.
❯ at the last line of the view as prompt. the plugin heavily use the Phantom API to process permission dialogs and diff review.
This approach shows that agent cli is highly viable in Sublime Text when combined with Phantoms, command palette, and EditView.
I have invented my own AI Shuttle Protocol ® where like Artemis II I depatch an XML container with payload to AIAgents (including Rules of Engagement) and have it returned ready to deploy immediately at Edge User after first scrutinising returned payload. I can inspect in Subl Text and prepare in Subl Text.
Some field data for this discussion: I recently implemented an agent-editor protocol client natively in ST — not ACP itself, but Claude Code’s IDE protocol (WebSocket + JSON-RPC + an MCP subset), which is a close architectural cousin: sublime-claude-code.
On the UI/API-limits concern upthread: for this integration surface, ST’s API turned out to be sufficient — phantoms give you accept/reject buttons on diff proposals, scratch buffers give you a side-by-side edit review, and the status bar covers connection state. The whole client is plugin-host Python 3.8, no dependencies, ~1,800 lines including a blocking diff-review flow. The main friction I hit was dev-time submodule hot-reload, not anything user-facing.
On the security concern: in this direction of integration the editor is the server — localhost-only, token-authenticated — and it acts as a UI/permission surface, not a tool executor. Nothing touches disk without an explicit accept in the editor.
FWIW, the ecosystem does seem to be converging on ACP (Claude Code already speaks ACP inside Zed via an official adapter). Having built this, I’d say a Sublime ACP client is very feasible — the protocol layer is genuinely thin.
Field notes from the other side, if useful: running six Claude Code sessions in parallel against one Sublime window, and what had to change to keep track of the output. https://dev.to/daichikudo/six-claude-code-sessions-in-parallel-and-only-reading-the-output-1i42
it’s already implemented: ACP client plugin for Sublime Text editor
i’m not the author, but it works fine for me
I will only add that I have regularly communicated between Subl and several AI Agents without plugins or packages. Cutting the tsunami of browser flow. Too much entropy in my mind. But that is my outlier view.
And I have made a first (according to Claude in writing out a certificate of innovation) in engineering moderator human (me) to discourse in a triad with two AI Agents. A “Meetings” environment.