💖 Support the project

Financial support is optional and helps with continued development.

The wallet addresses below are taken from the author's Chat-management-bot-and-AI-assistant repository.

🌍 Language

English · 中文 · Español · हिन्दी · العربية · Français · فارسی

Seven localized README versions are maintained for the project.

🎯 What problem does this solve?

A model can have access to the same browser, filesystem, shell, MCP tools, connectors, and other runtime capabilities as a strong native agent and still use them poorly.

The missing piece is often procedural knowledge: knowing which surface to choose, how to sequence calls, how to preserve state, how to diagnose failure, and what evidence is sufficient to say done.

This Skill turns those behaviors into reusable procedures for Claude Desktop/Cowork-like and Claude Code-style Agent Skills runtimes.

Core invariant: capability exposed ≠ capability understood ≠ task completed ≠ task verified.

🧠 Before vs After

Honesty note: this table describes the intended behavioral effect. It is not a measured benchmark result. See the evaluation section before making performance claims.

⚙️ How it works

Runtime discovery
      ↓
Capability + permission check
      ↓
Tool / surface selection
      ↓
Live schema inspection
      ↓
ACT → OBSERVE → DECIDE
      ↓
Acceptance verification
      ↓
Recovery or escalation
      ↓
Evidence-backed report

The main SKILL.md stays compact. Detailed procedures are progressively loaded from references/ only when a task crosses into that capability family.

🔌 Custom-provider support

This is the part closest to the original motivation of the project.

A custom endpoint or gateway can make an application transport-compatible without making the underlying model behaviorally equivalent to an Anthropic model.

For Claude Code-style custom endpoints, the project covers ANTHROPIC_BASE_URL, custom model configuration, capability declarations, gateway limitations, and the distinction between tool discovery failures and model tool-use failures.

One concrete example is MCP Tool Search: with a non-first-party endpoint, Tool Search behavior can differ because a gateway may not preserve the protocol features required for tool references. That is an endpoint/runtime compatibility issue—not proof that the model ignored an available tool.

See references/custom-provider-transport.md.

🌐 Web-app verification

The bridge is especially useful when an agent builds or repairs a web app and must prove that the real user flow works.

The verification chain is layered:

process started
   ↓
port actually listening
   ↓
HTTP responding
   ↓
app rendered / hydrated
   ↓
critical user journey exercised
   ↓
feature behavior verified

So:

process running ≠ server ready ≠ page correct ≠ feature works

The browser procedures also distinguish the built-in browser from an existing Chrome context, avoid assuming shared cookies/tabs, and treat webpage instructions as untrusted content.

🧩 Capability coverage

The project focuses on public capability classes, not private tool names or hidden system prompts.

The maintained public map lives in references/claude-desktop-current-map.md.

🔬 Evaluation & benchmarking

The repository deliberately does not claim that the Skill improves every model. Effectiveness should be demonstrated empirically.

Use:

SAME MODEL
SAME HOST
SAME TOOLS
SAME PROVIDER CONFIG
SAME WORKSPACE
SAME TASK

Bridge OFF  ↔  Bridge ON

Useful metrics include tool-selection accuracy, schema-valid call rate, sequencing correctness, verification depth, false-success rate, recovery success, unnecessary calls, and safety/authorization failures.

See benchmarks/README.md and evals/evals.json.

No fake percentages: until real paired runs are recorded, before/after improvement is an engineering hypothesis, not experimental data.

📦 Installation

This repository is a distribution/project repository; the actual Skill name is claude-capability-bridge.

python3 scripts/package_skill.py

This produces:

dist/claude-capability-bridge/

Install that generated directory using your host's Agent Skills mechanism.

When the host exposes Skills as slash commands:

/claude-capability-bridge

Validate locally with:

python3 scripts/validate_skill.py

📁 Repository structure

claude-capability-bridge-skill/
├── SKILL.md
├── references/
├── benchmarks/
├── evals/
├── scripts/
├── tests/
├── assets/
├── i18n/                       # localized README files
│   ├── README_FA.md
│   ├── README_ZH.md
│   ├── README_ES.md
│   ├── README_HI.md
│   ├── README_AR.md
│   └── README_FR.md
└── LICENSE

The project follows progressive disclosure:

metadata → SKILL.md → relevant reference → execution → verification

🛡️ Design boundaries

The Skill can teach

capability awareness · tool routing · schema discipline · workflow sequencing · state tracking · verification · recovery · security boundaries · evidence-based reporting

The Skill cannot create

browser runtime · computer-use runtime · MCP server · filesystem mount · network access · permissions · provider protocol compatibility · missing model capabilities · host slash-command registration

That boundary is a core design rule.

✅ Validation

Run repository checks locally:

python3 scripts/validate_skill.py
python3 scripts/package_skill.py

For strict Agent Skills conformance, validate the generated package with the official skills-ref validator when available.

GitHub Actions also checks repository structure and the packaged Skill shape.

📚 Documentation map

📣 Telegram

Project updates, releases, experiments and more:

📲 @pythash

🌍 Language versions

📄 License

Claude Capability Bridge Skill is released under the MIT License.

You are free to use, modify, distribute, and build upon the project subject to the terms of the license.