CCHarbour¶
A Claude Code-style agentic coding assistant — in your terminal, in a single ~2 MB binary that runs on Windows, macOS and Linux, written in Harbour.
╭─────────────────────────────────────────────────────────────────────────────╮ │ Welcome back, Anto! │ Tips for getting started │ │ ██████╗ ██████╗ │ Type a request to begin │ │ ██╔════╝██╔════╝ │ Run /help to list commands │ │ ██║ ██║ │ /caveman for terse replies │ │ ╚██████╗╚██████╗ │ ────────────────────────────────── │ │ ╚═════╝ ╚═════╝ │ What's new │ │ CCHarbour v0.8.18 │ background subagents + /tasks │ ╰─────────────────────────────────────────────────────────────────────────────╯ ╭─────────────────────────────────────────────────────────────────────────────╮ │ > how is the codebase organised? │ ╰─────────────────────────────────────────────────────────────────────────────╯ [superpowers] [tdd]
Why CCHarbour¶
🦾 Real agentic loop¶
Multi-iteration tool calls, streaming SSE, mid-turn interrupts, subagent dispatch. Stops only on a final answer or the iteration cap.
📦 One file, ~2 MB¶
A single console executable. No Python, no Node, no Docker. Drop it on a server, on a USB stick, in a CI runner.
🪟 Truly cross-platform¶
Windows (MSVC or mingw-w64), Linux (gcc), macOS (clang). The same Harbour source on every platform; a Win32 + POSIX console layer underneath.
🧠 Skills & Plan mode¶
Drop a Markdown checklist under .ccharbour/skills/ and the agent picks it up. /plan locks file writes until you approve the approach.
🛡 Permission gate¶
Every tool is allow, ask or deny. shell and edit prompt by default; you can flip any tool per project in settings.json.
🎯 Subagents with a user gate¶
propose_agents lets you review and approve batches of subagents before any of them runs. Each subagent has its own isolated context.
🪶 Lean mode¶
/lean trims the system prompt by ~500–800 tokens per turn for marathon sessions or pricier models.
✂️ Paste-aware input¶
Multi-line paste collapses to a tidy [pasted N lines text] placeholder. The full content is restored transparently on submit.
🧵 Background subagents¶
dispatch_agent_background spawns a subagent on a worker thread and returns a task-id (bg1, bg2, ...) immediately. Inspect with /tasks, /tasks view <id>, kill with /tasks kill <id>.
🎯 /goal — keep working until done¶
/goal <text> pins an objective and arms an auto-continue loop: the agent emits a GOAL COMPLETE sentinel when done, otherwise the REPL auto-feeds Continue toward the goal. between turns (cap 25).
💾 Full-state /save & /load¶
/save [name] snapshots messages plus goal, modes, skills, timer and the pending suggestion to JSON. /load round-trips the whole session, not just the chat.
🪝 Shell hooks¶
/hook add turn_complete <cmd> fires a user-defined shell command after every turn (success, error, interrupted). Fire-and-forget, six CCHARBOUR_* env vars (status / model / tokens / duration_ms / cwd / event), opt-in append-only log. Beep, toast, Slack webhook — pick your own side effect.
30-second tour¶
Once running, type a request. The agent reads, edits, runs commands and replies inline. See Getting started for the full walkthrough.
Common use cases¶
- Refactoring — explore a foreign codebase, plan the change with
/plan, watch the agent execute and diff every edit. - Code review — activate the
code-reviewskill, paste a diff, get a structured report with blocking and non-blocking findings. - Bug hunting — the
debuggingskill enforces reproduce → isolate → hypothesise → verify before any patch lands. - Project bootstrap —
/initwrites aCC.mdso future runs already know your conventions. - CI lint runner — pipe a prompt into
ccin non-interactive mode; permissions stay strict, output goes to logs.
Try it without installing¶
The web playground runs a JavaScript port of CCHarbour entirely in the browser — file, web, github, memory and todo tools — using your own DeepSeek API key. Great for kicking the tyres before downloading anything.
CCHarbour is released under the MIT licence.
Read the disclaimer before granting shell permission on a machine you care about.