SkillMesh

Stop stuffing hundreds of tools into your LLM prompt. Route to the right ones.

The Problem

LLM agents break when you load every tool into the prompt. Token counts explode, accuracy drops, and cost scales linearly with your catalog size. Teams with 50+ skills end up with bloated system prompts that confuse the model and burn budget.

SkillMesh solves this with retrieval-based routing: given a user query, it selects only the top-K most relevant expert cards and injects them into the prompt — keeping context small, accurate, and cheap.

Before vs After

	Without SkillMesh	With SkillMesh
Prompt tokens	~50,000+ (all tools loaded)	~3,000 (top-K only)
Tool selection	Model guesses from a huge list	BM25+Dense retrieval picks the best match
Cost per call	High (full catalog every time)	Low (only relevant cards)
Accuracy	Degrades as catalog grows	Stays consistent
Multi-domain tasks	Confusing for the model	Routed precisely (clean + train + deploy)

How It Works

User Query
    │
    ▼
┌─────────────────────┐
│  BM25 + Dense Index  │  ← Scores every card in your registry
└─────────┬───────────┘
          │
          ▼
┌─────────────────────┐
│   RRF Fusion Rank    │  ← Merges sparse + dense rankings
└─────────┬───────────┘
          │
          ▼
┌─────────────────────┐
│   Top-K Card Select  │  ← Returns the K best expert cards
└─────────┬───────────┘
          │
          ▼
┌─────────────────────┐
│  Agent acts as expert │  ← Full instructions injected into prompt
└─────────────────────┘

Each card contains: execution behavior, decision trees, anti-patterns, output contracts, and composability hints — everything the agent needs to act as a domain expert.

60-Second Demo

git clone https://github.com/varunreddy/SkillMesh.git
cd SkillMesh
pip install -e .
skillmesh emit \
  --provider claude \
  --registry examples/registry/tools.json \
  --query "clean messy sales data, train a baseline model, and generate charts" \
  --top-k 5

Output (truncated):

<context>
  <card id="data.data-cleaning" title="Data Cleaning and Validation Expert">
    # Data Cleaning and Validation Expert
    Specialist in detecting and correcting data quality issues...
  </card>
  <card id="ml.sklearn-modeling" title="Scikit-learn Modeling and Evaluation">
    ...
  </card>
  <card id="viz.matplotlib-seaborn" title="Visualization with Matplotlib and Seaborn">
    ...
  </card>
</context>

Only the relevant experts are injected — the rest of the 90+ card catalog stays out of the prompt.

Integrations

Platform	Method	Status	Docs
Claude Code	MCP server	Supported	Setup guide
Claude Desktop	MCP server	Supported	Setup guide
Codex	Skill bundle	Supported	Setup guide

Claude MCP Server

pip install -e .[mcp]
export SKILLMESH_REGISTRY=/path/to/tools.json
skillmesh-mcp

Exposes two tools via MCP:

• route_with_skillmesh(query, top_k) — provider-formatted context block
• retrieve_skillmesh_cards(query, top_k) — structured JSON payload

Copy-ready config templates in examples/mcp/.

Codex Skill Bundle

$skill-installer install https://github.com/varunreddy/SkillMesh/tree/main/skills/skillmesh

Quickstart

Install

python -m venv .venv && source .venv/bin/activate
pip install -e .[dev]

Optional extras:

pip install -e .[dense]   # Dense reranking with sentence-transformers
pip install -e .[mcp]     # Claude MCP server

Retrieve top-K cards

skillmesh retrieve \
  --registry examples/registry/tools.json \
  --query "set up nginx reverse proxy with SSL" \
  --top-k 3

Emit provider-ready context

skillmesh emit \
  --provider claude \
  --registry examples/registry/tools.json \
  --query "deploy container to GCP Cloud Run" \
  --top-k 5

Curated Registries

Use domain-specific registries for tighter routing:

Registry	Domain	Cards
tools.json / tools.yaml	Full catalog	90+
ml-engineering.registry.yaml	ML training & evaluation	15
data-engineering.registry.yaml	Pipelines & data platforms	10
bi-analytics.registry.yaml	BI & dashboards	10
devops.registry.yaml	DevOps & infrastructure	8
web-apis.registry.yaml	API design & patterns	7
cloud-gcp.registry.yaml	Google Cloud Platform	7
cloud-bi.registry.yaml	Cloud BI	5
roles.registry.yaml	Role orchestrators	10

skillmesh emit \
  --provider claude \
  --registry examples/registry/devops.registry.yaml \
  --query "configure prometheus alerting and grafana dashboards" \
  --top-k 3

CLI Commands

Command	Description
skillmesh retrieve	Top-K retrieval payload (JSON)
skillmesh emit	Provider-formatted context block
skillmesh index	Index registry into Chroma for persistent retrieval
skillmesh-mcp	Stdio MCP server for Claude

skillmesh --help

Repository Layout

src/skill_registry_rag/
├── models.py          # Tool/role card models
├── registry.py        # Registry loading + validation
├── retriever.py       # BM25 + optional dense retrieval
├── adapters/          # Provider formatters (codex, claude)
└── cli.py             # skillmesh CLI

examples/registry/
├── tools.json         # Full tool catalog
├── tools.yaml         # YAML version of full catalog
├── instructions/      # Expert instruction files (90+)
├── roles/             # Role orchestrator files
└── *.registry.yaml    # Domain-specific registries

skills/skillmesh/      # Codex-installable skill

Contributing

See CONTRIBUTING.md for how to add expert cards, create registries, and submit PRs.

Troubleshooting

skillmesh: command not found

pip install -e .

Missing registry path

export SKILLMESH_REGISTRY=/path/to/tools.json
# or pass --registry on every command

skillmesh-mcp fails to start

pip install -e .[mcp]

Codex does not detect new skill

Restart Codex after running $skill-installer.

Development

ruff check src tests
pytest

License

MIT — see LICENSE.

---

If SkillMesh helps your team, please star the repo — it directly improves discoverability and helps others find the project.