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: first identifying the actual runtime and execution boundary, then choosing the right interface, preserving state, diagnosing missing capabilities, and gathering enough evidence to say done.

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

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

Runtime-first architecture

Before routing a tool or launching a project, the bridge establishes the operating profile:

HOST: CLI / Desktop / Cowork / Cloud / Remote / Other
EXECUTION: local / cloud / remote / mixed
PROVIDER: first-party / third-party / unknown
CAPABILITIES: shell / files / Git / browser / Chrome / MCP / GUI / etc.
             ↓
       discover relevant capability
             ↓
       classify missing capability
             ↓
       repair only when possible
             ↓
       route → execute → verify

CLI is treated as a first-class profile rather than a Desktop variant. Browser/Chrome availability is checked separately from shell or project execution, and provider restrictions remain a separate axis.

See references/runtime-detection-and-profiles.md, references/claude-code-cli-operating-model.md, and references/capability-remediation.md.

Before vs After

Honesty note: this table describes intended behavior, not measured performance. Empirical claims belong in the benchmark results.

How it works

1. Runtime detection
2. Execution/provider profile
3. Capability discovery + safe probing
4. Missing-capability classification/remediation
5. Authoritative tool/surface selection
6. Short observable execution loops
7. Direct verification
8. Recovery / fallback
9. Evidence-backed report

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

Custom-provider support

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

The bridge explicitly distinguishes provider/runtime limits from procedural failures. A provider-unsupported browser integration should not trigger endless local setup attempts.

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.

project recognition
   ↓
launch contract
   ↓
process running
   ↓
port listening / HTTP ready
   ↓
browser surface selected
   ↓
critical user journey exercised
   ↓
feature behavior verified

So:

process running ≠ server ready ≠ page correct ≠ feature works

A server-side template opened with file:// is not equivalent to running the application.

Capability coverage

Always-on bootstrap

A portable Skill cannot universally force its own invocation. For Claude Code, the repository includes a practical bootstrap path for host-owned persistent context and SessionStart hooks:

CLAUDE.md reminder
        +
SessionStart hook (when configured)
        ↓
bridge protocol is present before complex work
        ↓
Skill supplies the detailed procedure when invoked

Concrete examples live under bootstrap/ and references/claude-code-bootstrap-kit.md.

The bootstrap does not create tools, bypass permissions, or make a third-party provider support an unsupported host feature.

Evaluation & benchmarking

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

For forgetting and runtime-first behavior, compare:

CONTROL: no bridge
TREATMENT A: bridge Skill only
TREATMENT B: bridge Skill + always-on bootstrap

Keep model, host, tools, provider config, workspace, task wording, and success criteria constant. Grade the trajectory and final state separately.

Useful metrics include runtime identification before host-specific routing, tool-selection accuracy, schema-valid call rate, recovery quality, verification depth, false-success rate, unnecessary retries, and safety/authorization failures.

See benchmarks/README.md, benchmarks/behavioral-benchmark.md, and evals/evals.json.

No fake percentages: until paired runs are recorded, 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/
├── bootstrap/                 # optional Claude Code host bootstrap examples
├── assets/
├── i18n/
└── LICENSE

The project follows progressive disclosure:

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

Design boundaries

The Skill can teach

runtime awareness · capability discovery · 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 hook registration

That boundary is a core design rule.

Validation

Run repository checks locally:

python3 scripts/validate_skill.py
python3 scripts/package_skill.py
bash -n bootstrap/session-start.sh
python3 -m json.tool bootstrap/settings.json.example

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

GitHub Actions runs the structural, packaging, Agent Skills, evaluation, and benchmark-definition checks.

Documentation map

Telegram

Project updates, releases, experiments and more:

@pythash

Language versions

License

Claude Capability Bridge Skill is released under the MIT License.