FonaTech/Clouds-Coder

Clouds Coder – A local‑first, web‑enabled runtime for AI‑driven coding agents

What it is

• A Python package (clouds‑coder) that runs a single process containing a CLI execution engine and a browser‑based IDE. The engine executes LLM‑driven task agents (e.g., code‑writing, debugging, refactoring) while the IDE shows the files, history, console output and a chat‑style “Agent collaboration” pane.
• The system is deliberately split into two planes:
  1. Execution plane – a local server that runs the LLM, tools, and the agent loop.
  2. User plane – a web UI (Monaco editor, terminals, debug console) that users interact with from a browser.
• It adds a governed LAN Collaboration Mode so multiple people (and their private agents) can work on the same project without the usual “last‑writer‑wins” problems.

Key capabilities

Feature	What it does
Session‑aware IDE	Isolated workspace per run, drag‑and‑drop file upload, full file‑system operations, undo/redo history, and a built‑in debugger (debugpy or pdb).
Agent collaboration UI	Chat view showing the agent’s plan, tool calls, and structured “Todo” cards; users can answer ask user prompts directly from the UI.
Prompt Enhancer	Optional “depth budget” (Low‑XHigh) that controls how far the model plans, explores alternatives, and validates results, without truncating the original request.
LAN Collaboration Mode	Separate listener (P+7 by default) that hosts a revision‑aware project workspace shared among approved devices. Handles presence, conflict resolution, and an immutable audit chain.
Write‑protocols	• Small UTF‑8 edits → operational‑transform (OT) with revision IDs. • Whole‑file/binary writes → optimistic‑concurrency (expected_revision). • Shell/agent processes → mutation lease that records before/after hashes and adopts writes atomically.
Security & Trust	Workspace‑declared MCP (model‑controlled‑process) commands stay inert until an admin explicitly approves the exact executable, args, env keys, and scripts. Any change revokes the approval.
Recovery‑oriented controller	Detects timeouts, truncations, and dead‑loop conditions; automatically creates continuation “Todo” tasks, injects hints, and retries with adjusted timing.
Admin plane	Web UI for project creation, device approval, password rotation, skill publishing, backup/quarantine, and a hash‑linked audit view.

Typical workflow

1. Start the runtime – python Clouds_Coder.py --host 0.0.0.0 --port 8080 --enable_collaboration.
2. Open the IDE at http://127.0.0.1:8085 (port + 5) and the Collaboration UI at http://127.0.0.1:8087 (port + 7).
3. Upload or create source files, then launch an LLM‑backed Agent from the UI (or via the CLI).
4. The agent writes to the workspace, posts its plan and evidence to the shared blackboard, and the UI shows live diagnostics.
5. If other team members join, they receive a revision‑aware view of the same files; conflicts are presented as candidate branches that must be reviewed before merging.

Who might use it

• Developer teams that want a reproducible, observable AI‑coding workflow without giving every user direct CLI access.
• Educators / labs teaching “Vibe Coding” (interactive AI‑assisted programming) where the UI can surface the agent’s reasoning for learning purposes.
• Tool builders who need a sandboxed environment to test LLM‑driven code‑generation pipelines with built‑in recovery and security checks.

Maturity

• Published on PyPI (version badge shown) with support for Python 3.8+.
• Multiple changelogs from May 2026 to August 2026 indicate active development, including major features like Skills Studio 2.0, structured truncation handling, and LAN collaboration security hardening.
• No mention of a stable 1.0 release yet, so the project is likely beta/early‑adopter level.

Installation

pip install clouds-coder

The package ships the Clouds_Coder.py entry point and the static web assets used by the IDE.

Getting started

1. Install the package.
2. Run the command shown above (adjust ports as needed).
3. Open the IDE in a browser, create a project, and invite teammates using the generated project password and device code.
4. Use the Prompt Enhancer dropdown to set the planning depth that matches your task complexity.

Security notes

• By default the collaboration service runs over plain HTTP and is intended for a trusted LAN. For any external exposure you must provide TLS cert/key (--collab_tls_cert/--collab_tls_key) or place the service behind a trusted reverse proxy.
• All MCP commands (commands that can execute arbitrary code on the server) are whitelisted only after admin approval, preventing accidental privilege escalation.

Where to look for more

• Detailed architecture and security changes are in the changelog files (log/CHANGELOG-2026-08-20.md, etc.).
• An example LLM configuration template is provided (LLM.config.json).
• The original inspiration is the shareAI‑lab/learn-claude-code project; Clouds Coder extends that kernel with the collaboration and recovery layers described above.

---

All information above is taken directly from the repository’s README; no external assumptions have been added.