A minimalist coding agent that works across your machines over SSH.

Run myco on your laptop. It edits files, runs shells, and searches code on the local machine and on every concrete Host alias in your ~/.ssh/config — one session, many hosts, no setup beyond SSH itself.

• One agent, many machines. Point tools at local or any Host alias from your ssh config (devbox, GPU box, CI host). Remotes attach over SSH on demand; you stay in a single conversation.
• Real computer use. Bash (including multi-turn sessions), a surgical file editor, text-mode browsing, and indexed search (keyword + semantic) on each host.
• Sessions you can resume. Titles, scratchpads, PR/worktree links, and full conversation history live under ~/.myco/ — pick up later with /resume.
• Nested agents for long work. myco drives itself: start myco in a bash session to spin off focused agents so the main thread stays small and cheap.
• Skills and project guidance stay searchable. Hosts auto-index skill packs and AGENTS.md / CLAUDE.md so the agent can find how you work.
• Coming later: multiplayer (multiple humans in the same agent workspace).

cargo install myco

Needs stable Rust, network on the first build (build.rs bakes MiniLM embedding weights into the binary via hf-hub), and ssh, lynx, uv, bash, tmux, fzf on PATH (git, gh, curl recommended).

myco    # runs the default model from your config.toml; --model <key> to switch
myco -p "explain build.rs"        # print mode: one turn, answer on stdout, exit
git diff | myco -p "review this"  # piped stdin becomes context for the prompt

-p/--print runs one non-interactive turn: the answer streams to stdout (raw, pipe-friendly), everything else prints to stderr, and the session is saved like any other (session=<id> on stderr) — continue it with --resume <id>. Bare -p takes the prompt from piped stdin.

Configure your models first: myco ships none built in. ~/.myco/config.toml holds a small catalog — [gateways.*] (protocol + base URL + auth, e.g. Anthropic, xAI, OpenRouter, or a local server) and [models.*] (the keys you pass to --model). The auth value is the token itself or a source such as { source = "env", var_name = "XAI_API_KEY" } (.env in the cwd is loaded at startup) or { source = "file", path = "~/.secrets/x.token" }. The exact variables are documented in the overview article — also available as myco --help overview once installed. Set a default model with model = "<id>" in ~/.myco/config.toml (--model wins). Transcript sections are colored when stdout is a TTY (--color auto|always|never; NO_COLOR / CLICOLOR_FORCE honored), and prose is word-wrapped with light markdown styling (--wrap auto|off|COLS caps the width at min(cap, terminal width), default 80; resizes reflow the transcript at the next prompt; never inside code blocks, never when piped).

Remotes just work: myco attaches lazily with ssh <alias> myco --mode host, so a remote only needs your key in ssh-agent and myco on the PATH used by non-interactive SSH. Runtime details: myco --help overview.

cargo test --locked --lib
cargo run --locked --bin myco
bash scripts/install-pre-commit-hooks.sh   # optional: CI bar (fmt + clippy) pre-commit

Semantic search embeds all-MiniLM-L6-v2 at compile time: build.rs downloads into the shared Hugging Face cache and bakes the weights into the binary — nothing large is in git. Offline seed: bash scripts/seed-minilm-weights.sh or MYCO_EMBED_CACHE.