agentic-signal

Enables users to construct AI agent workflows through a visual node-and-edge graph interface built on react-flow, where nodes represent discrete operations (LLM calls, data transforms, conditionals) and edges define execution flow. The platform serializes the visual graph into an executable workflow definition that can be interpreted by the runtime engine, supporting branching logic, loops, and multi-step orchestration without requiring code authoring.

Abstracts multiple local LLM providers (Ollama, Gemma, Llama) behind a unified interface, allowing workflows to invoke language models without cloud dependencies. The platform manages model loading, prompt formatting, and response parsing through a provider-agnostic adapter pattern, enabling users to swap between local models or providers by changing configuration without modifying workflow logic.

Enables building multi-step agent workflows where each step can invoke an LLM, process results, and pass outputs to subsequent steps. The platform orchestrates the execution sequence, managing context and state across steps. Supports agent patterns like chain-of-thought, tool use, and iterative refinement through workflow composition without requiring agent framework code.

Provides a library of pre-built node types (LLM inference, data transformation, conditionals, loops, API calls) that can be composed into workflows. Each node type encapsulates a specific operation with configurable inputs/outputs and execution semantics. The system supports custom node registration, allowing developers to extend the platform with domain-specific operations through a plugin-like mechanism without modifying core runtime.

Interprets serialized workflow graphs and executes them sequentially or in parallel depending on graph topology, managing state across node executions. The engine handles control flow (branching, loops), error propagation, and intermediate result caching. Execution occurs entirely locally without cloud orchestration services, with state persisted in-memory or to local storage depending on configuration.

Enforces a strict local-execution model where all workflow data, model inputs, and intermediate results remain on the user's machine. The platform does not transmit data to external APIs or cloud services by design, with no telemetry or analytics collection. This is achieved through exclusive use of local LLM runtimes and avoiding any cloud-dependent integrations in the core platform.

Published as open-source on GitHub with TypeScript implementation, enabling community contributions, auditing, and self-hosting. The codebase is structured for extensibility with clear separation between core runtime, UI components, and node implementations. Users can fork, modify, and deploy custom versions without licensing restrictions.

Serializes visual workflows to JSON format that captures node definitions, connections, and configurations. This enables workflows to be exported, version-controlled, shared, and imported across instances. The JSON schema is human-readable and can be manually edited or generated programmatically, supporting workflow-as-code patterns.

Provides node types for conditional branching (if/else logic based on runtime values) and loops (iterate over collections or repeat until condition met). These control flow nodes evaluate expressions at runtime and direct execution to different paths based on results. Supports nested conditionals and loops for complex workflow logic without requiring code authoring.

Provides visibility into workflow execution through step-by-step tracing, allowing users to inspect intermediate values at each node, view execution logs, and identify where failures occur. The debugger captures input/output for each node execution and supports pausing/resuming execution for interactive debugging without requiring code-level debugging tools.

Supports dynamic prompt construction through template variables that are substituted with runtime values from previous workflow nodes. Templates can reference node outputs using placeholder syntax (e.g., {{nodeId.outputField}}), enabling context-aware prompts that adapt to workflow data. Supports filters and transformations on substituted values for formatting.