Agent Client Protocol (ACP)

An open standard for connecting coding agents to clients (editors/IDEs), modelled on the Language Server Protocol (LSP) that decoupled language tooling from editors. If both agent and client speak ACP, any agent plugs into any editor. Covered in Full Course Spec-Driven Development with Coding Agents (DeepLearningAI):

• The ACP registry automates finding, installing and connecting agents inside a client — e.g. JetBrains' AI chat installs OpenCode (open-source agent) with one click, including the agent itself.
• Scope is broader than chat: it covers things like next-edit suggestions and plan mode; you can also write your own custom agent and install it locally.

The four-standard portability stack (per the course)

Standard	Layer
MCP	Agent ↔ external tools/data
AGENTS.md	Project rules for any agent
Agent Skills	Repeatable workflows + context (open standard; a skill ran unchanged in Codex after copying from Claude Code)
ACP	Agent ↔ editor/client

Thesis: models and agents change too fast (leaderboards churn), so keep your workflow — specs, skills, rules — above any one agent.

Why it matters for enterprise IT

A hedge against locking a whole engineering workforce into a single coding-agent vendor: standardise the IDE and the spec/skill assets, swap agents underneath. Worth noting the source's vendor interest — the course is a JetBrains partnership and JetBrains ships an ACP registry.

note Background (not from the source, unverified here): ACP was introduced by Zed Industries in 2025 and later adopted by JetBrains.

Related

CLI vs API vs MCP · Skills (Claude Code) · Spec-Driven Development (SDD) · Codex · Claude Code