Look at the release notes of almost any popular open-source agent on GitHub and two acronyms keep appearing: MCP and ACP. Together they point to a quieter but important trend, which is the standardisation of how agents talk to tools and to the programs that host them.
MCP: connecting agents to tools
The Model Context Protocol defines how an agent discovers and calls external tools and data sources. GitHub maintains its own MCP server, and its v1.8.0 release was tracked by automated dependency workflows in August 2026. Agents such as goose build their extension systems around MCP, so a tool server written once can serve many different clients.
ACP: connecting agents to interfaces
The Agent Client Protocol tackles the other side. It separates the agent runtime from the interface, so the same agent can sit behind a terminal, an editor, a chat bot or a web dashboard. By early June 2026 the specification had reached v0.13.5, and projects including Gemini CLI, Codex’s app-server and GitHub Copilot CLI were already being wrapped by adapters in a multi-agent plugin. Qwen Code graduated its --acp flag from an experimental name.
Neutral governance
The Agentic AI Foundation under the Linux Foundation, which counts Anthropic, OpenAI and Block among its backers, now hosts goose and provides a neutral home for this kind of shared infrastructure. Neutral governance matters because companies are more willing to build on standards that no single vendor controls.
What developers should do
If you build agent tooling, support MCP for tools and consider ACP for editor integration. If you consume agents, prefer projects that implement open protocols, since they are easier to swap out. Protocols are still changing quickly, so pin versions and read changelogs before upgrading.