Goose Roadmap 2026 Q3

[DOsinga]

🪿 Goose Roadmap 2026 Q3

This is the first roadmap where the focus of the project is to provide a developer kit to improve the ease by which developers can build their own agents. If you have a new feature, check whether it fits in with our general direction and then discuss on Discord.

🔌 ACP as the goose Client Protocol

Our desktop app talks to the goose binary over ACP(+): standard ACP where applicable, unstable _goose prefixed routes where not. We'll audit the _goose routes and resolve each into one of three buckets: migrate those that duplicate existing ACP functionality onto the standard methods we should have used from the start; propose the ones with broad value (like session renaming/archiving and mid-session extension changes) as official additions to the spec; and explicitly mark the rest as goose-specific: product features like recipes, schedules, and dictation. We'll also clean up the protocol where needed, then publish and document it so that developers looking to integrate with goose have the easiest path possible.

📚 Goose as a library

We’ve published the providers crate as a reusable component with UniFFI bindings. We’ll organize the rest of the goose system into a set of useful crates covering all functionality. Currently we support Python and Kotlin as supported languages, we’ll add Swift and more based on demand.
A key part of this work is unrolling the agent loop into a re-entrant state machine that advances one turn at a time, with the conversation itself as the state. The loop will instead be composed from independent operations for model calls, tool approval and execution, compaction, elicitation, retries, subagents, and other behaviors. This will make the agent loop easier to test, embed in event-driven systems, and customize: developers will be able to replace individual operations, add application-specific policies, change behavior mid-session, and integrate goose with orchestration systems for long-running or horizontally scaled agents.

🧭 2026 MCP spec

The new MCP spec changes the core protocol in a number of ways, with stateless operation and better discovery support being the most impactful. We’ll help make rmcp, the Rust SDK for MCP, handle the new protocol cleanly, then use that work to keep Goose’s MCP client, built-in servers, and extension flows aligned with the standard.

🛠️ Goose powertools

The primary future of Goose is as a developer kit: the GDK. We will continue supporting the desktop app as a reference client, with developers as its primary audience. To support that role, we’ll turn it into a workbench for agentic development, with tools for inspecting conversations, model interactions, tool calls, context construction, hooks, token usage, latency, cost, and caching. We’ll also add import and export for common agent trace formats, edit-and-replay workflows for debugging, and better controls and visibility for context management.