OpenMobius-skill: ICT/SMC Trading Knowledge for AI Coding Agents

What OpenMobius-skill Does and Who It Is For

OpenMobius-skill addresses the gap between general-purpose AI coding agents and domain-specific trading analysis. The standard response from an AI coding agent to a trading question draws on training data, which may conflate concepts from different methodologies or produce analysis that sounds plausible but is not grounded in any specific framework. OpenMobius-skill installs a local knowledge base of ICT/SMC concepts, configures the agent to route trading questions through scoped knowledge retrieval, and adds current market data and chart generation for supported analysis modes.

The primary audience is traders and quantitative developers who use Claude Code, Codex, OpenClaw, Hermes, Cursor, or WorkBuddy for their development workflow and also practice ICT or Smart Money Concepts analysis. The skill installs locally and requires no subscription beyond the user's existing agent API access.

The Knowledge Base and School/Source Routing System

The core of the skill is 2,008 knowledge cards organized into Schools. ICT/SMC is the default School and handles standard concepts like Liquidity Sweep, Order Block, and Fair Value Gap. Other Schools include ChanLun (a Chinese market analysis methodology) and an SMC material source called Wuyuan. Each School's knowledge is stored as separate vectors in ChromaDB.

The routing system resolves both the user's intent and the analysis route before any knowledge retrieval occurs. A request with no explicit School selector defaults to strict ICT/SMC. An explicit School selector like "using ChanLun only" hard-filters retrieval to that School and never silently falls back to ICT/SMC. The README gives a concrete example: asking about 中枢 using ChanLun only retrieves attributable ChanLun knowledge without mixing in ICT terminology.

The capability gate is a distinct step in the pipeline. Before fetching market data or generating a chart, the skill checks whether the requested analysis has a native market analyzer. ChanLun Q&A works without market data, but requesting a ChanLun market analysis fails the gate and returns a capability gap response rather than silently relabeling an SMC analysis as ChanLun.

Installing OpenMobius-skill

The installation uses a temporary clone pattern. The README recommends creating a temporary directory, cloning the repository into it, running the installer, and then removing the clone:

OPENMOBIUS_SRC="$(mktemp -d "${TMPDIR:-/tmp}/openmobius-src.XXXXXX")"
git clone https://github.com/MobiusQuant/OpenMobius-skill.git "$OPENMOBIUS_SRC"
cd "$OPENMOBIUS_SRC"
python3 install.py --platform claude-code
cd "${TMPDIR:-/tmp}"
rm -rf -- "$OPENMOBIUS_SRC"

The installer copies source files into `~/.claude/skills/openmobius-skill/` for Claude Code (other platforms have their own skills directories). From there it creates a `.venv/`, installs dependencies, downloads Playwright's Chromium browser (approximately 280 MB into the OS user-global cache), downloads the `nomic-embed-text-v1.5` embedding weights (approximately 547 MB into the HuggingFace cache), builds the vector index, and runs a health check.

On Windows, use `py -3 install.py --platform claude-code` or the provided `install.ps1` PowerShell script. The `--platform all` flag installs for all five supported local-path hosts in one run; WorkBuddy uses local ZIP import instead.

After Install: Capability Discovery and Analysis Modes

After installation, the skill registers itself with the platform's skill loading system. The README provides example invocations for testing the install:

"What analysis models can I use?"
"What is Liquidity Sweep?"
"How is ETH 4h looking?"

The first example triggers capability discovery, which reads the installed registry dynamically and reports native market analysis profiles, Q&A-only lenses, knowledge categories, and available composition modes without fetching market data. This distinguishes between Schools that support live analysis (ICT/SMC has a built-in market analyzer) and Schools that are Q&A only (ChanLun at default install has no native market analyzer).

Three composition modes are available: strict (only the selected School's evidence controls the analysis), augment (one School controls the bias and trade levels while another provides labeled supporting context), and compare (separate attributed analysis branches run in parallel and conflicting definitions stay separate). The README gives an example of the augment mode: 'Use SMC as primary and Wuyuan as a reference to analyze BTC 1h' runs SMC for the primary signal while Wuyuan evidence appears with explicit attribution.

Chart Generation and Market Data Handling

Chart generation uses Playwright with the lightweight-charts library. When a supported analysis route passes the capability gate, the skill generates a visual chart unless the user opts out. For user-supplied OHLCV data pasted as CSV, the skill preserves that snapshot and extracts structure locally rather than replacing the user's data with a live series.

If a user attaches a chart image and asks for analysis, the skill attempts to identify the asset and timeframe from the visible data. If it cannot identify them reliably, the result is labeled as visual-only with the limitation disclosed rather than fabricating identifiers.

Real-time market data fetches only apply when a route passes the capability gate, the user's request includes a named asset and timeframe, and the user does not attach a CSV snapshot. Pass-through to the indicator API requires the user to name a specific indicator; the skill does not auto-fetch indicators that the user did not request.

Limitations and Dependency Footprint

The skill installs approximately 827 MB of cached data across the Playwright Chromium download and the Nomic embedding model. This is a significant footprint for a skill that runs inside an existing coding agent. Teams with limited disk space or corporate proxies that block large downloads should verify connectivity before installing.

The current-market analysis feature requires a native analyzer, and not all Schools have one. The README states that ChanLun requests for market analysis stop before market-data or chart work rather than producing a mislabeled result. This is a deliberate design choice, but it means the skill's analytical depth is uneven across methodologies.

A comparable alternative is asking an AI coding agent to fetch market data and apply technical analysis using the `ta` Python library or similar. That approach gives more flexibility in indicator choice but provides no grounding in specific trading methodologies like ICT's liquidity concepts. OpenMobius-skill's differentiation is that knowledge retrieval is attributable: each answer cites which School and source produced the evidence, making it possible to trace a concept back to its methodology.

License, Platform Support, and Maintenance

OpenMobius-skill is licensed under Apache-2.0. The last push was on 2026-09-04. The repository has no GitHub releases; the version is tracked in the source files. The CHANGELOG.md file records changes between versions.

Supported platforms at time of the last push are Claude Code, Codex, OpenClaw, Hermes, Cursor, and WorkBuddy. WorkBuddy requires a local ZIP import rather than the CLI installer. The `--platform all` flag covers all five local-path hosts in one invocation.

Runtime dependencies include chromadb>=0.5, sentence-transformers>=5.3, transformers>=5.5, numpy>=1.24, einops>=0.7, Pillow>=10, playwright>=1.40, and optionally openai>=1.50 for remote OpenAI-compatible embedding endpoints. The full list is in requirements.txt.

Editorial conclusion

OpenMobius-skill fits quantitative developers and traders who use AI coding agents for strategy development and want grounded ICT or SMC knowledge without hallucinated analysis. The first install is time-intensive: roughly 280 MB for Playwright's Chromium browser and 547 MB for the Nomic embedding model download. Verify that the target platform is in the supported list (Claude Code, Codex, OpenClaw, Hermes, Cursor, or WorkBuddy) before installing, and run `python3 install.py --platform <name>` from a temporary clone of the repository rather than a permanent location, since the clone is a one-shot source bundle that the installer removes after use.

Frequently asked questions

Official sources

1. Issues
2. License: Apache-2.0
3. MobiusQuant/OpenMobius-skill on GitHub
4. Project website
5. README