D'AUBE // BRIDGE²

Write once. Run across agents.

D'AUBE // BRIDGE² is an open-source universal AI skill compiler and MCP bridge. Define one portable capability spec, then compile it into integration artifacts for ChatGPT/Codex, Claude, Gemini, DeepSeek, Meta/Llama, MCP hosts, and Chromium browsers.

External review

On 2026-09-18, the BRIDGE² deployment record was AI Deploy Network Approved after review of the submitted implementation evidence. This means the deployment record passed AI Deploy Network review; it is not a client-confirmed commercial outcome or a guarantee of production performance.

The canonical publication narrative is maintained in docs/AI_DEPLOY_CASE_STUDY.md. The reviewed case study is now publicly available on AI Deploy Network.

Why BRIDGE²

AI tooling is fragmented. The same capability is repeatedly rewritten as MCP tools, provider function schemas, skills, plugins, and browser configuration. BRIDGE² treats those surfaces as compiler targets while keeping capability logic provider-neutral.

skill.yaml
  |
  +-- MCP tool schema
  +-- OpenAI / ChatGPT / Codex artifacts
  +-- Claude plugin + SKILL.md
  +-- Gemini SKILL.md + MCP config
  +-- DeepSeek tool schemas
  +-- Meta/Llama function schemas
  +-- Chromium extension manifest

Current release: 7 target families, 15 deterministic artifacts, one MCP server.

30-second start

git clone https://github.com/daubesonntag-dotcom/daube-agent-bridge.git
cd daube-agent-bridge
python -m pip install -e ".[dev]"
daube-bridge validate examples/skill.yaml
daube-bridge compile examples/skill.yaml -o build/sample

Install the verified v0.1.3 wheel directly from GitHub Releases:

python -m pip install https://github.com/daubesonntag-dotcom/daube-agent-bridge/releases/download/v0.1.3/daube_agent_bridge-0.1.3-py3-none-any.whl
python -m daube_bridge doctor

python -m daube_bridge ... is the shim-free CLI path for Windows environments where Application Control blocks generated console-script executables.

Run BRIDGE² as a local stdio MCP server:

python -m daube_bridge.server

Or expose the HTTP MCP endpoint for remote/local integrations:

daube-bridge serve
# http://127.0.0.1:8000/mcp

One-click MCPB

BRIDGE² ships an MCPB v0.4 bundle definition using the cross-platform uv runtime. Build the deterministic bundle with:

python scripts/build_mcpb.py

The build is reproducible: identical source produces an identical archive hash. Official MCP Registry metadata lives in server.json.

Target matrix

This matrix describes integration artifacts BRIDGE² generates. It does not invent provider capabilities that do not exist.

Portable skill format

name: My Research Skill
version: 0.1.0
description: Research and compare options.
tools:
  - name: compare_options
    description: Compare options against explicit criteria.
    parameters:
      type: object
      properties:
        options:
          type: array
          items: {type: string}
      required: [options]

MCP tools

The bundled server exposes three small primitives:

• bridge_targets — discover supported compilation targets.
• validate_skill — validate a portable skill object.
• compile_skill — compile a skill in-memory for another agent or tool.

That makes BRIDGE² usable both as a CLI and as infrastructure callable by other agents.

Verify before release

python scripts/verify.py
python scripts/build_mcpb.py
.tools/mcp-publisher.exe validate

The verification gate covers tests, Ruff, target discovery, sample compilation, wheel creation, deterministic MCPB creation, and official MCP Registry metadata validation.

After publishing a GitHub Release, smoke-test the public wheel in an isolated environment:

python scripts/smoke_release.py 0.1.3

Roadmap

• ChatGPT Apps SDK interactive widget target.
• npx / TypeScript compiler parity.
• Provider capability detection and compatibility linting.
• Signed bundles and provenance manifests.
• Remote registry + searchable skill catalog.
• One-click adapters for agent IDEs and orchestration frameworks.
• Golden interoperability fixtures across providers.

Project principles

1. Portable by default. Prefer MCP and open schemas over provider lock-in.
2. Evidence over pretending. Never claim native support that a provider does not have.
3. Readable outputs. Generated artifacts stay human-auditable and diffable.
4. Safe boundaries. Authentication and sensitive actions remain explicit.
5. Fast adoption. A useful first result should take minutes, not a framework migration.

Contributing

Issues, adapters, compatibility fixtures, docs, and provider integrations are welcome. Provider-format changes should include a link to authoritative documentation and a minimal reproducible fixture.

Useful starting points:

• CONTRIBUTING.md
• CHANGELOG.md
• GOVERNANCE.md
• docs/PROVIDER_COMPATIBILITY.md
• docs/20K_STAR_PLAYBOOK.md
• docs/CODEX_FOR_OSS.md
• docs/OPEN_SOURCE_FUND.md
• docs/DISTRIBUTION.md
• docs/PERFORMANCE_EVIDENCE.md
• docs/THREAT_MODEL.md
• docs/AI_DEPLOY_NETWORK.md
• docs/LAUNCH_KIT.md
• SECURITY.md

License

MIT © 2026 D'AUBE SONNTAG.