Muse Code
Muse Code is Meta’s terminal coding agent, powered by Muse Spark 1.2, for handling complex software engineering tasks across large repositories. The agent can plan changes, write code, validate results, and coordinate multiple persistent subagents during development sessions. Muse Code uses async background agents that stay active throughout a session to reduce repeated information gathering and help complete multi-step tasks with less steering. Its runtime uses a local event log that records model calls, tool runs, approvals, and edits so sessions can be replayed and resumed after failures. Muse Code includes bundled skills such as /plan for approval-gated planning, /grill for stress-testing plans, and /goal for working toward completion. Built for AI developers and software teams, Muse Code helps automate coding workflows, long-running engineering tasks, debugging, and repository-level development.
Learn more
eve
eve is an open source framework for building AI agents that you own, designed around a file-system-based structure and production-ready runtime. An instructions.md file is enough to define an agent, while optional building blocks such as model configuration, skills, tools, sandboxes, channels, connections, subagents, schedules, and evals can be added as the agent grows. Tools are created by dropping TypeScript files into the tools directory, and skills are written as Markdown playbooks that load only when relevant. Agents can connect to services such as GitHub, Stripe, and Linear through authenticated connections and can surface through Slack, Discord, Teams, web chat, APIs, cron jobs, CLIs, and custom applications from the same codebase. Specialized subagents can handle delegated work with their own prompts, tools, and sandboxes.
Learn more
Flue
Flue is an open agent framework for building durable AI agents with a programmable TypeScript harness. Created by the team behind Astro, it uses a React-like hooks API to compose agent behavior, persistent state, lifecycle events, models, tools, sandboxes, subagents, skills, and MCP servers in code. Agents are stateful and addressable over HTTP, keep context across conversations and events, and can change capabilities as work progresses. Flue records every session in a durable stream, so accepted work survives crashes, restarts, and deploys; interrupted sessions resume automatically and clients reconnect without starting over. Developers can run agents locally, in CI, from their own backend, or orchestrate them with systems such as Cloudflare Workflows and Inngest. Secure sandboxes let agents run commands, edit files, and do real work, while tools connect them to APIs and data. Flue is powered by Pi and supports different LLM providers, allowing teams to choose the models they want.
Learn more
Muse Spark 1.1
Muse Spark 1.1 is a multimodal reasoning model from Meta Superintelligence Labs built for agentic tasks, coding, computer use, tool use, and multimodal understanding. The model improves on the original Muse Spark with stronger performance in planning, orchestration, long-context work, coding workflows, and external app interactions. Muse Spark 1.1 can manage a 1 million token context window, remember earlier actions, retrieve important information, compact context, and delegate tasks across parallel subagents. It is designed to operate across tools, MCP servers, custom skills, browsers, native apps, scripts, images, video, PDFs, and audio-based workflows. Developers can access Muse Spark 1.1 through the new Meta Model API public preview, while users can try it in Thinking mode in the Meta AI app and on meta.ai.
Learn more
