Axel

A terminal-first workbench for coding agents, built on Rex.
Website: axel.md · docs: docs.axel.md.

$ axel            # inside Rex: the launcher floats over the window.
                  #   why not today? — the prompt. Type what you want, Enter starts it
                  #                    with the agent shown at the right (the last one
                  #                    used, or the one an @name picks), in place, here.
                  #                    Tab opens the full form (projects, agent, worktree).
                  #   threads        — recent threads below it: ↓ to get there, Enter opens
                  #                    one, Ctrl+S settles, Ctrl+X archives, Ctrl+D deletes.
                  #   usage          — a small HUD of every agent: state, tokens, sparkline.
$ axel            # again: closes them (toggle)

Threads are the primitive: an intent, attached to any number of projects, worked by many
agent sessions over time on any machine, shared with other people. Every change is a signed
event, so replicas merge. Agents are named Ed25519 identities with an inbox. Their status flows back through
Claude Code hooks, Codex notify, ACP, or axel status, never by scraping a terminal,
and is mirrored to Rex through OSC 7501 so Rex's tab indicators agree with Axel.

See docs/DESIGN.md. Install: just install (release build, symlinked into ~/.local/bin and ~/.cargo/bin).

Commands


`axel`	floating launcher / inbox (inline when not in Rex)
`axel ui [--screen launch\|inbox]`	what the floating pane runs
`axel agent new NAME [--harness claude\|codex\|opencode] [--model M] [--skills a,b\|*]`	create an agent (key, attestation, record)
`axel agent ls\|show\|archive`
`axel thread new "intent" --agent NAME [--project DIR]... [--scratch] [--worktree [BRANCH]]`	thread + first session
`axel thread spawn THREAD --agent NAME`	another agent on the same thread
`axel thread resume THREAD [--agent NAME]`	new session continuing the last harness conversation
`axel thread attach THREAD --project DIR`
`axel thread open THREAD`	the thread view: conversation, folded tool calls, composer
`axel thread ls [--all]`, `show`, `stop`, `done`, `archive`
`axel session host SESSION`	hosts a managed (ACP) session; its pane is the raw log
`axel session ls [--all]`, `axel session stop SESSION`
`axel log SESSION [-f]`	tail a managed session's raw frames
`axel send "@atlas please review the diff"`	signed message into an agent's inbox
`axel inbox [--agent NAME]`
`axel status working\|blocked\|done\|… "msg"`	status from scripts (reads `$AXEL_THREAD`)
`axel hook claude\|codex`	hook sink, installed automatically at launch
`axel doctor`

Runtimes

managed (default when the harness's ACP adapter is on PATH): Axel drives the agent over
the Agent Client Protocol. You work in the thread view: your messages, the agent's prose,
questions inline; tool bursts fold into one line (Space expands); l jumps to the host pane,
which prints the raw frames. Permissions are decided by Axel's policy, never by a dialog.
Adapters: npm i -g @zed-industries/claude-code-acp (Claude), codex-acp (Codex), goose acp.
pane: the harness's own CLI in a Rex pane; status via hooks.

Autopilot

Agents default to autopilot: Claude Code runs with --permission-mode bypassPermissions
(Codex with --full-auto) and Axel's guard hook still denies irreversible shell commands.
Axel pre-accepts Claude Code's folder-trust dialog for the project you picked. Claude Code's
bypass permissions warning is yours to accept once per machine: either press
"Yes, I accept" the first time, or set it in ~/.claude/settings.json:

{ "skipDangerousModePermissionPrompt": true }

(The older bypassPermissionsModeAccepted key in ~/.claude.json is only read by a
one-time migration into that setting.)

Rex key binding

Append docs/rex-init.lua to ~/.config/rex/init.lua, then rex config reload.

Axel

Commands

Runtimes

Autopilot

Rex key binding

Reviews (0)