• 

  Safe workspace start: Open an isolated browser copy to protect real project files while human and agent inspect shared state.

• 

  Proof of Completion: Deterministic test validation and cryptographic SHA-256 proof with instant one-click rollback.

• 

  Review First policy: Direct file writes are strictly blocked until a human inspects the diff and grants explicit approval.

• 

  Exact Diff Preview: WebMCP agent generates atomic code and config changes without modifying actual disk files.

• 

  Evidence Engine & Lock: Ranks relevant project assets with confidence scores and locks hashes to prevent stale drift.

• 

  DELTA: Model-neutral shared WebMCP workspace bringing evidence, controlled autonomy, and rollback to humans and agents.

DELTA — The Shared Evidence Workspace for Humans and AI Agents

"Don't give your agent the whole project. Give it DELTA."

💡 Inspiration

Today's coding agents are remarkably capable, yet they frequently suffer from critical real-world failure modes:

1. Context waste & hallucination: Dumping entire repositories into an LLM context window burns tokens, misses cross-file dependencies, and invites hallucinations.
2. The "Trust me, it's done" dilemma: Agents often declare a task finished without providing deterministic proof or a verifiable diff.
3. Irreversible mistakes: When an agent modifies the wrong file or corrupts a configuration, reverting changes requires tedious manual Git archaeology.

We asked: What if humans and AI agents worked inside the exact same visible, state-aware workspace governed by the Model Context Protocol (WebMCP)? This inspired DELTA—a model-neutral, local-first shared workspace where agents receive bounded evidence, propose atomic diffs, prove completion, and provide instant rollback.

⚙️ What It Does

DELTA turns complex, mixed software projects into precise, state-aware WebMCP tools with a fail-closed human-in-the-loop workflow:

• Universal Safe Workspace: Users open a safe in-browser project copy. Original files remain completely untouched during exploration and planning.
• Evidence Engine & Evidence Lock: Rather than sending thousands of lines of code, DELTA ranks relevant files, symbols, imports, tests, and documentation with calibrated confidence. The resulting evidence set is hashed into an Evidence Lock; if files drift, execution halts automatically.
• Exact Diff Preview: External AI agents call WebMCP tools (plan_task, preview_change) to generate atomic replacements, additions, or deletions without touching the disk. Previews render immediately in the human Review changes view.
• Human-Owned Autonomy: Under Review First policy, direct file mutations are strictly blocked until a human inspects the diff and grants one-time approval. Agents cannot elevate their own permissions or forge approval.
• Proof of Completion & 1-Click Rollback: Tasks are verified against deterministic success criteria. Semantic success generates a cryptographic SHA-256 Completion Proof. If anything is unsatisfactory, clicking Rollback restores the original state in under a second.

🛠️ How We Built It

DELTA was engineered from the ground up as a high-performance, local-first web application:

• WebMCP Architecture: Deep implementation of the standard WebMCP specification via document.modelContext.registerTool() and executeTool(). Over 41 dynamic tool contracts adapt in real-time to the open project state (e.g., expanding from 6 discovery tools to 22 task-specific tools as a project and diff are formed).
• Local-First Indexing & Graph Analysis: Fast in-memory symbol, AST-like textual references, and dependency graph traversal built in modern modular JavaScript/ES modules.
• Zero External LLM Lock-in: DELTA does not embed an LLM; it is purely the evidence, boundary, and execution layer that any WebMCP-compatible host (OpenAI Codex, ChatGPT, Claude, local agents) can seamlessly command.
• E2E Automation & Verification: Comprehensive regression suite powered by Playwright and Node.js test runners ensuring 100% deterministic reproducibility across browser sessions.

🥊 Challenges We Faced

• Dynamic Tool Registry Synchronization: Ensuring that WebMCP tool availability matches the exact current UI state (e.g., apply_change becoming available only when a preview exists, and bind_agent_native_host only when requested by a human) without race conditions.
• Balancing Safety and Agent Autonomy: Creating a system where agents feel natural and flexible in proposing multi-file refactors while guaranteeing that unauthorized file writes, binary corruption, and path traversal remain mathematically impossible.
• Designing Trustworthy Completion Proofs: Moving beyond "LLM says OK" to deterministic criterion evaluation ($\text{SHA-256}(\text{evidence} + \text{diff} + \text{tests})$) that both humans and automated evaluators can audit.

🏆 Accomplishments & Key Learnings

• True Model Neutrality: Demonstrated that the WebMCP layer is the operational interface, allowing any agent to perform surgical, multi-file code and configuration changes with zero secret leakage.
• Bounded Context Efficiency: Proved that feeding an agent a calibrated 10–20 KiB evidence slice yields higher task accuracy than dumping a 200 KiB codebase into the prompt.
• Reversible Peace of Mind: Building an instant rollback engine that restores complex multi-file projects back to baseline with a single click.

🚀 What's Next for DELTA

• Expanded Adapter Ecosystem: Introducing declarative adapter SDKs for game engines (Godot, Unreal, Unity) and complex domain formats.
• Multi-Agent Handoff Protocol: Broadening our integrity-checked task handoff schema so heterogeneous agents can collaborate on long-horizon engineering tasks without passing privileged credentials.

Built With

• access
• agent
• agents
• ai
• api
• automation
• bridge
• codex
• css
• evidence-based
• file
• html
• javascript
• native
• node.js
• openai
• project
• python
• rollback
• security
• system
• unity
• webmcp