Terminal coding agent. Single binary ccb. Talks to Anthropic (OAuth or API key), Anthropic-compatible endpoints, ChatGPT Codex (OAuth), OpenAI-compatible endpoints (Ollama, DeepSeek, vLLM, …), and Gemini through one agent loop. macOS, Linux, Windows.
Solo-maintained public derivative of Claude Code; baseline reconstructed from the v2.1.88 sourcemap (2026-03-31). See ATTRIBUTION.md. Not affiliated with Anthropic.
macOS / Linux (also Git Bash / WSL):
curl -fsSL https://raw.githubusercontent.com/Icarus603/claude-code/main/install.sh | bashWindows (PowerShell):
irm https://raw.githubusercontent.com/Icarus603/claude-code/main/install.ps1 | iexNo Node, no Bun, no package manager required. Auto-updates on every startup.
Open a terminal — on macOS press ⌘ Space, type Terminal, hit Enter. On Linux press Ctrl Alt T. On Windows open PowerShell from the Start menu.
Then run:
ccbThat drops you into an interactive REPL. Type any task in plain English and press Enter:
> refactor the auth module to use async/await
> explain what this codebase does
> write tests for src/utils/parser.ts
One-shot mode (no REPL, prints and exits):
ccb -p "what does package.json do?"Continue the last conversation:
ccb --continueInside the REPL:
| Key | Action |
|---|---|
Enter |
Send message |
Shift+Enter |
New line |
Escape |
Cancel current response |
/help |
Show all slash commands |
Ctrl+C |
Exit |
Windows ARM64 is not supported. Use x64 under emulation, or WSL.
New here? Inside the REPL, type:
/powerup
Ten short interactive lessons walk you through what makes ccb worth using — @ file mentions, permission modes, /rewind and Esc-Esc to undo, background sessions, CLAUDE.md, MCP, skills + hooks, /fork for parallel branches, model dialing, and multi-provider switching. Open one, try it, mark it done. The home banner tracks progress until you've finished all ten.
ccb keeps a persistent, file-based memory under ~/.claude/projects/<project>/memory/ — a MEMORY.md index plus typed topic files (user, feedback, project, reference). It writes to these as it works so a fresh session can orient fast. Over many sessions those files drift: entries go stale, duplicate each other, or contradict the current code. /dream is the reflective pass that cleans them up.
Inside the REPL, type:
/dream
It runs right now, in the foreground, with full tool access while you watch. The model does a four-phase pass over your memory directory:
- Orient —
lsthe memory dir, readMEMORY.md, skim topic files and recent session logs so it improves rather than duplicates. - Gather — pull new signal from session logs and transcripts (grepped narrowly, never read whole), and flag memories that have drifted from the code.
- Consolidate — merge new signal into existing files, convert relative dates to absolute, delete contradicted facts.
- Prune & index — keep
MEMORY.mdan index (one line per entry, under 200 lines / ~25 KB), drop superseded pointers, reconcile againstCLAUDE.md.
When it's done you get a short summary of what was consolidated, updated, or pruned.
| Command | What it does |
|---|---|
/dream |
Consolidate now — foreground, full tools, you're watching. Accepts trailing text as extra context (e.g. /dream focus on the auth refactor). |
/dream nightly |
Schedule a recurring nightly consolidation. Installs a durable, recurring cron that fires /dream consolidate at a random minute between 00:00–05:59 local (jittered to avoid thundering-herd across sessions). The schedule persists in .claude/scheduled_tasks.json across sessions. |
/dream consolidate |
The bare consolidation body with no manual-mode preface — what the nightly cron fires. You can also run it by hand. |
/dream also has the alias /learn.
- Recurring jobs auto-expire after 7 days — re-run
/dream nightlyto renew. (Re-running also de-dupes: it deletes any existing/dream consolidatejob before scheduling a fresh one.) - Cancel anytime — list jobs and delete by ID via the cron tools, or toggle the Auto-dream row in
/memory. - Requires auto-memory to be on. The command is hidden in remote mode and when auto-memory is disabled.
/loop turns any prompt or slash command into a repeating job. Two ways to run it: a fixed interval, or no interval at all — in which case the model decides how long to wait between iterations based on what it just saw.
Inside the REPL:
/loop 5m /babysit-prs # run /babysit-prs every 5 minutes
/loop 30m check the deploy # run a plain prompt every 30 minutes
/loop check the deploy # no interval → model self-paces the cadence
/loop # bare → autonomous check, dynamically paced
The interval is the leading token (5m, 2h, 1d) or a trailing every … clause (check the deploy every 20m, run tests every 5 minutes). Granularity is 1 minute minimum. /loop runs the task immediately, then schedules the next firing — you don't wait for the first tick.
| Form | Behaviour |
|---|---|
/loop <interval> <prompt> |
Fixed cadence. Converts the interval to a cron, fires recurring until you cancel. |
/loop <prompt> (no interval) |
Dynamic mode — after each run the model picks the next delay (via ScheduleWakeup) from what it observed: quiet branch → wait longer, lots in flight → wait shorter. |
/loop (bare) |
Autonomous default in dynamic-pacing mode — runs a check now, self-paces from there. |
/loop also has the alias /proactive.
Notes
- Recurring (fixed-interval) jobs auto-expire after 7 days — re-run to renew. Dynamic-mode loops stop the moment the model declines to schedule the next wake-up.
- List jobs with
/cron-list, cancel with/cron-delete <id>. - If an interval can't be expressed cleanly as cron (e.g.
7m,90m), the model rounds to the nearest clean cadence and tells you what it picked. - GA by default. Disable the whole scheduler locally with
CLAUDE_CODE_DISABLE_CRON=1.
A TUI dashboard for orchestrating background sessions. Type:
ccb agentsYou get a live list of every background session grouped by state (Needs input · Working · Done), a dispatch input to launch new ones, and a peek panel for reading any session's recent activity without attaching.
Inside the agents view:
| Key | Action |
|---|---|
Type then Enter |
Dispatch a new background session with that task |
Shift+Enter |
Newline in the dispatch input |
↑ / ↓ |
Move focus between sessions |
→ |
Attach into the focused session |
Space |
Peek the focused session (reply inline without attaching) |
Tab |
Toggle the agents drawer / accept suggestion |
@name / /cmd |
Mention an agent, skill, or repo |
Shift+↑ / Shift+↓ |
Reorder within a bucket |
Ctrl+R |
Rename the focused session |
Ctrl+T |
Pin the focused session to the top |
Ctrl+X |
Stop / delete the focused session (two-step) |
Ctrl+S |
Switch grouping (by state ↔ by directory) |
Ctrl+G |
Open the dispatch buffer in $EDITOR |
| Click | Click a row to focus it; click inside the input to move the cursor |
? |
Open the in-view help overlay |
Esc |
Clear input, then exit |
Ctrl+C |
Confirm-then-exit (background sessions keep running) |
Sessions are PTY-backed and survive after you close the terminal — re-open ccb agents later to find them. Dispatch goes through a spare-worker pool so a fresh session boots in ~0ms when one is ready.
Agent Teams let you coordinate several ccb instances working together. One session acts as the team lead — it assigns work, synthesizes results, and coordinates. Each teammate is a full, independent session with its own context window, and teammates message each other directly. Unlike a plain subagent (which runs inside one session and only reports back to the main agent), you can also talk to any teammate directly without going through the lead.
You have to ask for a team — it won't form on its own. Describe the task and the team shape in plain English and the lead sets everything up. ccb may also propose a team if it spots parallelizable work, but it always waits for your confirmation first. Either way, nothing spawns without your say-so.
> Users report the app exits after one message instead of staying connected.
Create an agent team: spawn 4 teammates to investigate different hypotheses,
have them message each other to disprove each other's theories like a
scientific debate, and update the findings doc with the consensus.
The lead creates a shared task list, spawns teammates, lets them claim and work tasks (with dependencies auto-unblocking as blockers complete), synthesizes the results, and cleans up the team when done. In the REPL, cycle teammates with Shift+Down and message any one directly; or run split panes (tmux / iTerm2) to watch everyone at once.
Teams shine when parallel exploration adds real value and teammates can work independently:
- Research & review — split a PR review or a library investigation into independent lenses (security, performance, test coverage) that run at once.
- New modules or features — each teammate owns a separate piece without stepping on the others.
- Debugging with competing hypotheses — teammates test different theories in parallel and argue them out, which beats a single agent that anchors on the first plausible cause.
- Cross-layer coordination — a change spanning frontend, backend, and tests, each layer owned by a different teammate.
When not to: teams add coordination overhead and burn significantly more tokens (each teammate is a separate session). For sequential work, same-file edits, or heavily dependent tasks, a single session or a subagent is the better tool. New to teams? Start with research or review — clear boundaries, no parallel-write conflicts.
- Start with 3–5 teammates. Past that, coordination overhead and token cost outweigh the parallelism; three focused teammates beat five scattered ones.
- Split along file boundaries. Two teammates editing the same file overwrite each other — give each its own set of files.
- Give each teammate enough context. They load
CLAUDE.md, MCP, and skills like any session, but they do not inherit the lead's conversation — put task-specific detail in the spawn prompt. - Steer as it runs. Press Enter on a teammate to read its session, Escape to interrupt; redirect approaches that aren't working rather than letting the team run unattended.
- GA by default in ccb — no env var, no CLI flag (upstream still gates it behind
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS). Teams are cleaned up at session end.
Requires Bun ≥ 1.3.
git clone https://github.com/Icarus603/claude-code.git
cd claude-code
bun install
bun run dev # hot-reload REPL
bun test
bun run doctor:archdoctor:archandbun testmust pass. No--no-verify.- Commit messages explain why.
- No npm publishing — binary-only by design.
- PRs are reviewed by one maintainer at human pace. No SLA.
- Open an Issue before non-trivial PRs.
# macOS / Linux
rm -rf ~/.local/share/ccb ~/.local/bin/ccb
# Windows (PowerShell)
Remove-Item -Recurse -Force "$env:LOCALAPPDATA\Programs\ccb"