Agentic Browser Control Plane (ABCP — Personal R&D)

Personal R&D. Not a product.
ABCP is an engineering playground for building a standards-first control plane that observes, explains, and governs agentic browser actions across Gemini Computer Use, AWS Bedrock AgentCore, and ChatGPT Atlas overlay surfaces.

Vision

1. Explainability you can hand to Audit — OpenTelemetry spans, structured policies, and signed evidence bundles.
2. Interoperability that survives platform churn — MCP-native drivers, modular tool adapters, and langgraph orchestration.
3. Reproducibility — Golden traces from BrowserGym/WebArena, Playwright trace viewer artifacts, and deterministic env controls.

Repo Layout

apps/
  control-plane/   # LangGraph runtime, drivers, policy hooks
  ui/              # Next.js 15 app, Storybook, accessibility tooling
packages/
  types/           # Shared Zod schemas and TypeScript contracts
  otel/            # OpenTelemetry GenAI helpers
  mcp/             # MCP clients + sample servers
  evidence/        # C2PA embeds, cosign signing, Rekor proofs
eval/              # BrowserGym/WebArena harness + golden traces
docs/              # Docusaurus documentation site
.github/workflows/ # CI/CD pipelines (lint, tests, a11y, provenance)

Quickstart

pnpm install
pnpm dev

• apps/control-plane: pnpm --filter control-plane dev
• apps/ui: pnpm --filter ui dev
• docs: pnpm --filter docs docs:dev

REST API Highlights

• POST /runs — create a new agent run (supply driver)
• POST /runs/:runId/actions — submit an action event for execution
• GET /runs/:runId/summary — aggregate run metadata, policy decisions, observations, and evidence bundle pointers in a single response

Evaluation Harnesses

Run reproducible BrowserGym/WebArena suites via uv:

# BrowserGym
cd eval/browsergym
uv sync
uv run python -m eval_runner tasks/sample_login.json

# WebArena
cd ../webarena
uv sync
uv run python -m eval_runner --suite tasks/sample_suite.json

Ethos & Guardrails

• Prime directives: correctness & tests → security & provenance → accessibility & UX → maintainability → performance.
• Observational Atlas driver: captures DOM/network evidence only; no introspection hooks.
• Compliance toggles: demonstrate California SB 243 (Companion Chatbot) and AI transparency disclosures.
• Provenance: Every evidence pack includes C2PA manifests, cosign signatures, and Rekor inclusion proofs.
• Accessibility: Storybook a11y, keyboard-first navigation, and WCAG 2.2 automation in CI.

Contributing

See CONTRIBUTING.md, CODE_OF_CONDUCT.md, and SECURITY.md.
All commits follow Conventional Commits and ship with tests, spans, and provenance updates.

Citation

If you build on this work for research, cite using the metadata in CITATION.cff.