Ciel

What is Ciel?
Ciel is an AI-powered platform that transforms natural language into production-ready Chainlink CRE workflows and lets users compose them into multi-step pipelines.
Users describe what they want automated on-chain — for example:
“Monitor ETH price every 5 minutes and swap on Uniswap when it drops below $1800.”
Ciel’s AI engine generates a fully compilable CRE TypeScript workflow, validates it against 14 safety constraints, simulates it via the CRE CLI, and publishes it to an on-chain registry on Base Sepolia.
But Ciel goes beyond single workflows. Published workflows become composable building blocks. Users can chain them into DAG pipelines through a visual drag-and-drop canvas — connecting a price monitoring workflow to a compliance check to a DEX swap — with automatic schema compatibility checking, conditional branching, parallel execution, and aggregated pricing. Other users and AI agents can discover, pay for, and execute individual workflows or entire pipelines via x402 micropayments.
The result is a three-layer flywheel: users generate workflows from natural language, publish them to an on-chain marketplace, and AI agents consume them — either individually or as composed pipelines. More workflows generated means a richer marketplace; a richer marketplace attracts more agents; more agent consumption generates more revenue for creators, attracting more builders.

How It Works
Ciel follows a 5-stage AI pipeline that converts user intent into a deployed workflow.

1. Intent Parsing (Deterministic NLP) - This stage performs zero-latency keyword extraction using Porter stemming and three-tier fuzzy matching (exact → stemmed → Levenshtein). It also performs trigger classification (cron, HTTP, or EVM log), entity recognition (e.g., “Uniswap” → dex-swap), and negation detection. No LLM calls are made here — it is purely algorithmic NLP running in sub-millisecond latency.
2. Template Matching (Hybrid Scoring) - The parsed intent is scored against 22 pre-built CRE templates covering DeFi, institutional finance, risk/compliance, and AI-powered categories. The system combines TF-IDF keyword scoring with semantic embeddings using ONNX all-MiniLM-L6-v2. This ensures the LLM customizes a known working scaffold instead of hallucinating code from scratch.
3. Code Generation (GPT-5.3-Codex with Structured Outputs) - Ciel generates complete CRE TypeScript using OpenAI’s Responses API with a layered prompt architecture. A static base prompt contains the CRE SDK reference and 14 critical constraints, while a dynamic layer injects template patterns and examples. The model performs a structured self-review before returning code to catch issues like async handlers or non-deterministic operations.
4. Validation and Auto-Repair - Generated code first goes through inexpensive regex validation checks, such as import whitelisting, async detection, and export verification. If those pass, the code is compiled using TypeScript. A deterministic quickFix system automatically repairs common LLM mistakes, such as replacing Date.now() with runtime.now(), updating JSON parsing for CRE v1.2.0 compatibility, and correcting transaction patterns.
5. Simulation and Deployment - The workflow is compiled from TypeScript to JavaScript and then to Javy WASM using the CRE CLI. Simulations run in an isolated workspace. When a user publishes a workflow, they sign an on-chain transaction to the AutopilotRegistry contract on Base Sepolia, and the backend triggers deployment to the Chainlink DON.

The full pipeline includes retry logic, template fallback, abort signals, compile-feedback loops, and an error resolver that enriches retry attempts with known fix patterns across multiple error categories.
Pipeline Composability — Chaining Workflows
Many real-world automations require multiple coordinated steps. Ciel’s Pipeline Composer lets users combine workflows into multi-step DAG pipelines.
Visual DAG Builder
Users drag workflows onto a canvas and connect outputs to inputs. The interface supports undo/redo, zoom, pan, and grid snapping. Connections are validated in real time.
Schema Compatibility Engine
Before connecting workflows, the system verifies that outputs match inputs using exact field matching, fuzzy matching, and runtime type coercion. If mismatches occur, the system provides diagnostics and auto-mapping suggestions.
Conditional Branching
Each pipeline step supports conditions such as eq, neq, gt, lt, gte, lte, and contains. Branching logic allows steps to route to different downstream workflows depending on results.
DAG Execution Engine
Workflows at the same stage run in parallel. Each step has a timeout and retry policy, and pipeline execution progress streams in real time using server-sent events.
Automatic Pricing and Execution History
Pipeline costs are calculated as the sum of the individual workflow execution prices. Each pipeline run records full execution traces and metrics.
Agent Auto-Composition
A reference AI agent can automatically compose pipelines from a goal string using compatibility scoring, price ranking, and reliability metrics.
Workflow Marketplace and Discoverability
Ciel includes a marketplace where workflows can be discovered, executed, and monetized.
Marketplace UI
Users browse workflows by category, chain, or keyword. Each workflow card displays capabilities, price per execution, reliability score, and creator address. Users can inspect code, view configuration, review simulation traces, and execute workflows directly.
On-Chain Publishing Flow
When a workflow is published, the user signs a transaction to the AutopilotRegistry contract on Base Sepolia using MetaMask. The backend confirms the transaction, updates the platform database, emits real-time events, and deploys the workflow to the Chainlink DON.
x402 Micropayment Gating
Each workflow execution endpoint is protected by x402 micropayments. Non-owners pay per execution using EIP-712 authorization, while creators can execute their own workflows without charge.
Real-Time Agent Activity Feed
A live event feed displays workflow generations, publications, executions, and pipeline runs in real time, providing visibility into platform activity.
Workflow Detail View
Each workflow has a dedicated page where users can inspect code, configuration, and simulation traces. Owners can redeploy, push to GitHub, or download the project.

How Agents Access Workflows
Ciel workflows function as payable microservices and can be accessed in multiple ways.

1. MCP Server -Agents such as Claude, GPT, Cursor, and Windsurf can access workflows through MCP tools for discovery, retrieval, and execution.
2. Buyer SDK -Developers can use the @ciel/sdk npm package to discover workflows, retrieve metadata, and execute them programmatically.
3. REST API -Agents can directly call API endpoints to discover workflows and execute them with x402 payment authorization.
4. On-Chain Registry - Agents can read workflow metadata directly from the AutopilotRegistry contract on Base Sepolia.
5. A2A Agent Card - Ciel publishes a standard agent-card endpoint describing its capabilities for automatic discovery by agent orchestration frameworks.
6. CLI - Command-line tools allow users to search, inspect, and execute workflows directly from the terminal.

Flagship Template — Multi-AI Consensus Oracle
The flagship template generates workflows that query multiple AI models (such as GPT-4o, Claude, and Gemini) on each Chainlink DON node.
Each node performs a local majority vote across models. The Chainlink DON then performs Byzantine Fault Tolerant consensus across nodes, producing a cryptographically verifiable answer that is written on-chain.
This architecture provides provable multi-party verification of AI outputs — something centralized systems cannot achieve.

What Problem It Solves
CRE is the most powerful decentralized compute platform available. It runs arbitrary TypeScript on the Chainlink DON with BFT consensus, cross-chain execution via CCIP, and verifiable access to external data. However, four critical gaps prevent CRE from reaching its full potential.
The Builder Gap — CRE is powerful but inaccessible.
Chainlink’s own data shows CRE reduced workflow development from one week to three hours. But three hours is still too long for someone who simply wants to automate an on-chain task.
Developers must understand WASM compilation constraints such as no async/await, no Date.now(), and no fetch. They must also understand the synchronous .result() pattern, the two-step report()/writeReport() flow, Javy QuickJS limitations, consensus-safe timestamp handling, and multiple SDK-specific APIs before writing a single workflow. Even though the CRE SDK documentation is strong, the cognitive overhead prevents most developers from getting started.
Ciel removes this barrier completely.
A user who has never used the CRE SDK can describe their automation goal in plain English and receive production-grade, simulation-tested, DON-deployable code in under 60 seconds. The platform uses 22 template scaffolds so the LLM customizes known-working code instead of generating everything from scratch. A 10-step quickFix auto-repair system catches common LLM mistakes when generating CRE code, such as replacing Date.now() with runtime.now(), correcting JSON.parse(body) to decodeJson(body), and fixing incorrect transaction patterns.
Chainlink reduced workflow development from weeks to hours. Ciel reduces it from hours to minutes.
The Reusability Gap — Every CRE workflow is a one-off.
Today, CRE workflows are isolated implementations. There is no standardized way to discover, share, or monetize them.
If a developer builds a workflow that monitors the ETH price and rebalances a portfolio, others who want the same automation must rebuild it themselves. The same work is repeated across the ecosystem.
Ciel introduces a permissionless workflow marketplace where published workflows are registered on-chain through the AutopilotRegistry contract on Base Sepolia. Workflows become discoverable through multiple integration paths including MCP, SDK, REST API, on-chain registry queries, A2A agent cards, SSE event streams, pipelines, CLI tools, and reference agents.
Anyone can discover and execute a workflow by paying for it. Each execution generates revenue for the workflow creator.
This creates a compounding ecosystem flywheel: more builders produce more workflows, which attracts more consumers, which generates more revenue and encourages more builders to participate.
The Composability Gap — Single workflows cannot express complex automations.
Real-world blockchain automation rarely fits into a single workflow. For example, a typical automation might require checking a price feed, verifying compliance conditions, executing a swap, and sending an alert.
Without composability, users must manually chain workflows together or write large monolithic codebases.
Ciel introduces a pipeline composer that allows users to visually chain workflows into DAG pipelines. These pipelines support conditional branching using seven operators, schema compatibility checking using a three-tier system (exact match, fuzzy match, and runtime type coercion), parallel execution, and aggregated pricing.
A reference AI agent can even autonomously compose pipelines from a goal description.
Instead of isolated scripts, the ecosystem becomes a composable network where each workflow functions like a reusable building block.
The AI Verification Gap — Single-model AI outputs are unreliable for on-chain decisions.
AI models are not perfectly reliable and can produce hallucinations. Empirical testing from Chainlink showed approximately 89.3% accuracy across categories for real Polymarket outcomes, dropping to about 84.3% for political questions.
For high-value on-chain financial decisions, relying on a single AI model introduces a dangerous single point of failure.
The corporate actions pilot conducted with Swift, DTCC, Euroclear, and multiple financial institutions demonstrated that multi-model consensus significantly improves accuracy for structured data. However, this architecture currently lacks an accessible developer implementation.
Ciel solves this with its Multi-AI Consensus Oracle template.
This template generates CRE workflows where each Chainlink DON node independently queries multiple AI models such as GPT-4o, Claude, and Gemini. Each node computes a local majority vote across models, and the DON then performs BFT consensus across nodes to produce a cryptographically verified answer written on-chain.
This two-layer verification structure is uniquely suited for decentralized infrastructure. A centralized server cannot prove that multiple independent machines verified agreement across several AI models.

Ciel makes this powerful architecture accessible through a single natural language prompt.
The core insight is that the main bottleneck in blockchain automation is not execution capability. The bottleneck is who can write workflows, who can discover them, and who can compose them.
Ciel transforms CRE from an expert-only SDK into an accessible platform. It turns individual workflows into composable, tradeable economic primitives and unlocks decentralized multi-model AI consensus as a capability any developer can generate in seconds.

How Is It Built?
Technology Stack
Ciel is built using a modern full-stack architecture designed to support AI workflow generation, decentralized execution, and real-time composable automation.
Monorepo and Build System
The project uses a monorepo architecture built with Bun workspaces and Turborepo. This setup enables parallel builds, shared packages, and efficient dependency management across backend, frontend, SDK, and agent components.
Backend Infrastructure
The backend is built using Express.js running on the Bun runtime. Data persistence uses SQLite with Drizzle ORM. The backend includes structured error handling through an AppError system, request rate limiting, and Server-Sent Events (SSE) to stream real-time updates to the frontend.
Frontend Application
The frontend is built with Next.js 14 using the App Router and React 18. The UI uses TailwindCSS together with shadcn/ui components. The Monaco Editor is used to display generated code. State management is handled with Zustand, while wallet connectivity is implemented using wagmi and RainbowKit.
AI Engine
The AI generation system uses OpenAI GPT-5.3-Codex through the Responses API with structured outputs for deterministic code generation. Semantic template matching is implemented using ONNX all-MiniLM-L6-v2 embeddings, while intent parsing uses a deterministic NLP layer with a Porter stemmer.
Pipeline Engine
Ciel includes a custom DAG pipeline execution engine. The engine supports conditional branching using seven operators, schema compatibility checks using a three-tier matching system (Levenshtein similarity and runtime type coercion), and parallel execution of steps within the same position group. Execution progress is streamed to the UI using SSE. Pipelines are composed visually through a drag-and-drop canvas interface.
Workflow Marketplace
The marketplace layer connects the application to on-chain infrastructure. Workflows are registered on-chain using the AutopilotRegistry and AutopilotConsumer contracts. Execution endpoints are protected using x402 micropayment gating. The marketplace supports category, chain, and keyword search, and provides a real-time activity feed powered by SSE. Each workflow has a dedicated detail page with code, configuration, and simulation views.
Smart Contracts
The on-chain contracts are written in Solidity 0.8.24 using the Foundry toolchain. The contracts are deployed on Base Sepolia and include AutopilotRegistry and AutopilotConsumer for workflow registration and execution reporting.
CRE Integration
Ciel integrates with the Chainlink CRE ecosystem using the @chainlink/cre-sdk version 1.1.3. The CRE CLI is used for compilation and simulation, converting workflows from TypeScript to JavaScript and then to Javy WASM.
Testing Infrastructure
The project includes a comprehensive test suite with 889 tests across 38 files using bun:test. This includes 137 tests covering the intent parsing system across all 22 templates and 76 pipeline tests across 7 files validating DAG execution and branching behavior.
Payments
Workflow executions are monetized using the x402 SDK. Payments use EIP-712 typed-data authorization and are settled in USDC through the CDP Facilitator.
Virtual TestNet Environment
Testing against realistic blockchain conditions is supported through Tenderly integration. This allows the platform to run forked mainnet-state simulations with snapshot and revert functionality.