an AI harness on zsh · ghostty · neovim · claude code
Every Claude session lives in a Neovim buffer, next to your shell and your dev server. Jump between them with the key combos you already use, and send any selection to Claude with your question or instruction.
One frame on an Astro site: Claude, a shell and astro dev as buffers, and a selection carried from the code to Claude.
vim trick
The Claude session is a buffer, so Vim marks work in it. When Claude gives you a table, a plan or a diff you'll want again, drop a mark on it. Then keep the conversation going as long as you like and jump straight back to that reference point.
Uppercase marks are global: set mT on the table and `T brings you back to it from any file in the frame.
# in the claude buffer, normal mode (Ctrl-\ Ctrl-n) ma mark the answer with the good table i back to the prompt: keep talking … twenty turns later … `a jump straight back to the table mT / `T uppercase: works from any buffer
resume
If most of your work with Claude happens in Claude Code sessions, the expensive part isn't the code. It's the context: the decisions, the dead ends, the "we tried that, it broke X".
frame shell --recent lists your latest sessions by their first message. Pick one and it resumes in the directory it ran in. Give Claude the latest update and decide where to go next.
--find "WORDS" searches what you typed, for when you remember what you said but not when.● marks a session that's already open. Picking it switches to that window instead of forking it.--resume ID when you already know the session.coordinate
Start in a shell frame: a Claude in any directory, with no branch and no worktree, so by design it isn't writing code. It plans. When change needs to happen, it frame spawn wts a worktree frame in whichever frame-enabled repo needs it, sends the work over frame req, and collects the replies in its inbox.
Each worker is fully isolated: its own branch, its own Neovim, its own ports and dev servers. The coordinator keeps the big picture; the workers keep their hands off each other.
spawn_frames=true in frame kv. With it off, the coordinator asks before spawning. macOSframe ls and sends the work there instead.frame wt TOPIC --from SRC hands a whole conversation forward to a new frame when the work changes shape.isolate
Running more than one Claude Code session on a repo usually means two dev servers fighting over a port, an agent editing main, and worktrees nobody remembers to clean up. A frame is self-contained and disposable, and delivery back to main goes through guarded steps.
A branch, a worktree, and assigned ports for its services. Agents work on their topic branch, never in the primary checkout.
Refuses on a dirty primary, fast-forwards to origin first, and aborts cleanly on conflict instead of leaving a half-merge.
Removes the worktree and branch, stops the dev servers it started, and reaps frame state. Nothing left running.
autonomy
frame kv tells each frame's Claude what it may do without asking. Set it per frame, per project, or for every project. Settings can change mid-session, and agents re-check before they act.
| autocommit | commit each finished, working step |
| merge_on_commit | then frame merge it into main |
| push_on_merge | push main to origin after merging |
| deploy_on_merge | and deploy when the project knows how |
| spawn_frames | spawn worker frames in other repos without asking |
get started
To get started with frame today, head to the repo. The Getting Started guide covers prerequisites, install, and onboarding your first project, from frame init to your first frame wt.
Built on zsh, Neovim and Claude Code, with Ghostty recommended. macOS is the primary target; the core workflow also runs on Linux and is tested there in CI.