One index per workspace

Every agent that opens a codebase reads it again from scratch. It walks the tree, greps for a symbol, opens the files around it, and throws the result away when the session ends. The next agent, on the next machine, repeats it.

mux builds that reading once. A workspace connects GitHub, chooses repositories, and mux keeps one index of them: a mirror per repository, split into syntax-aware chunks, and on paid plans embedded so a search can rank by meaning rather than by words. Everyone in the workspace searches the same index — from the Home page, from Fluxcode after a device signs in, and from any agent over MCP.

What a search returns

Three ways in, and all of them return the same shape: the repository, the file path, the start and end lines, a score, an excerpt, and the commit the excerpt was read at.

• grep_code matches the shape of the syntax tree. foo($X) finds every call of foo with one argument, whatever that argument is — not a regular expression, and not a text match.
• search_code ranks by meaning, so a search for “retry a failed upload” can return code that never uses those words.
• ask_code takes a question in plain words and returns the passages that bear on it. It does not write the answer; your agent reads the passages and does that.

corpus_status and list_repositories complete the set. Five tools, no workspace argument: the key names the workspace.

What the plans draw

Pattern search is on every plan, because it costs a process and a tree, not an embedding. Hobby indexes one repository and refreshes it when you ask. Pro indexes five and re-indexes on push. Team indexes every repository the GitHub installation grants and re-indexes on push. Search by meaning and ask_code come with Pro and Team; on Hobby they refuse in plain words rather than quietly returning something else.

What the index does not do

It does not hold your code: the repositories stay in GitHub, and deselecting one removes its mirror and its part of the index. It reads only what the installation grants. And it does not answer for you — the index returns code with line numbers, and the model you already trust does the rest.