loom

织流 Loom is a locally-run agent pipeline workbench designed to solve the full chain of complex tasks such as mathematical modeling, research reports, and long-document writing. It breaks tasks into editable multi-step workflows, where each step can bind a skill, artifact filename, role (executor/reviewer/editor), engine and model, and supports setting checkpoints to pause and await confirmation. Core features include: - Orchestration: Visually add/remove steps, flexibly configure each step. - Skill Library: Manage prompt specifications for each step, supporting creation, editing, save-as-copy, and importing standard skill packages (zip / SKILL.md). - Execution: Real-time streaming via SSE, the transcript area shows tool calls and output step by step, the process card in the upper right corner corresponds to each step's status, and the workspace file panel lists files the agent is currently writing in real time (markdown rendered in place, images/PDFs previewed inline). - Conversational Local Revision: When unsatisfied after a run, you can point to a specific step and say how to change it, rewriting only that step without rerunning the entire workflow. - Workflow Import/Export: One workflow = one JSON file, shareable, backup-able, and re-importable. One hard rule: the body content can only be produced by the local CLI agent. Loom itself does not directly connect to any large model API to generate content; the API presets in settings are only used to inject endpoints/keys/models into Claude Code or Codex CLI. Before use, you need to install and log into one of them: Claude Code (the claude command available) or Codex CLI (the codex command available). Running from source: pip install -r requirements.txt python run.py # http://127.0.0.1:8000 python run.py --port 8001 # change port python run.py --auto-kill # auto-release port occupied by old process Regression tests (no network required): python -m pytest -q Directory structure: - app/: FastAPI backend, including main.py (routes and all /api), runner.py (workflow execution engine: checkpoints, SSE event streams, workspace, artifacts), agents.py (Claude / Codex CLI adapter layer), updater.py (check for updates, download installer, verify sha256). - static/: Frontend (vanilla HTML/CSS/JS, no framework, no build step). - tests/: pytest regression tests, including static asset contract tests. - loom_launch.py: Packaged entry point (pywebview native window, falls back to browser on failure). - run.py: Source-mode entry point. - make_release.py: PyInstaller onedir + Inno Setup to release/Loom-[version]-setup.exe. - upload_cos.py: Upload installer and latest.json to Tencent Cloud COS. Data directory: Source mode runs in modex-data/, installed mode runs in data/ at the same level as the exe (SQLite + workspace + custom skills + exports). Upgrades do not overwrite it, and uninstallation does not delete it. Packaging and release: python make_icon.py # only needed if icon changed python make_release.py --version 1.0.1 --notes "..." # produces setup.exe + latest.json python upload_cos.py # upload to COS (requires local key file, not in repo) The installer only includes the application itself and dependencies; it does not package the development machine's modex-data/ (make_release.py has an explicit check for this). Privacy: No server-side account system, no telemetry, no uploading of any content. Workflows, prompts, and artifacts all stay local. The only external requests are reading the version manifest during update checks, and the agent CLI's own communication with model services. Download: https://lxlrwxs.top/modelflow/ | Requires Windows 10/11 | Fully local execution, no account, no telemetry.