Open-source agentic framework

Model-driven reasoning.
Structured execution.

Loomspan composes LLM-backed skills and Java services into a hierarchy of tasks. Models decide how to approach a mission; your application defines the capabilities and boundaries.

Start with the framework Explore the execution model

JAVA 21+ / SPRING BOOT / APACHE 2.0

A MISSION, DECOMPOSEDILLUSTRATIVE
APPLICATION INPUT

“Plan a trip around my preferences.”

planTripYAML
  • understandPreferencesYAML
  • planTransportYAML
    • searchFlights / searchTrainsJAVA
    • rankTransportOptionsJAVA
  • planStayYAML
    • searchHotelsJAVA
  • assembleItineraryYAML
Model-backed skill Java capability
01
FrameworkPlan and execute inside your application
02
ConsoleInspect the runtime and its traces
03
Agent integrationBring runtime evidence into development

01 / Execution model

A task hierarchy.
Built as the work unfolds.

Loomspan combines Hierarchical Task Network (HTN) structure with LLM-driven decomposition. A planning skill selects from its allowed children; each specialist can work within a smaller, local set of capabilities.

01

Declare the capabilities

Define reasoning in YAML and deterministic operations in Java. Both register as named skills in one catalog.

YAML + @SkillMethod
02

Let the model decompose

A planner forms tasks from the mission and the skills it may call. Nested skills handle their own part of the work.

allowed_skills + max_steps
03

Follow the execution

Tasks run through a shared lifecycle. Events and finalized traces provide evidence of decisions, calls, results, and failures.

Events + execution traces

02 / Authoring skills

Reason in YAML.
Act through Java.

A skill is a named capability with an explicit contract. It can reason with a model, delegate to other skills, or call an ordinary Spring service.

In the travel sample, planTransport can choose a flight or train search and ask Java to rank the results. The parent receives a structured transport digest.

Read the complete travel sample
The sample uses fixture inventory. The tree above shows selected capabilities, not a fixed execution sequence.
plan_transport.ymlEXCERPT 
name: planTransport
description: >
  Transport specialist planner.
model: qwen3-35b
planning_mode: true
max_steps: 6

allowed_skills:
  - name: searchFlights
  - name: searchTrains
  - name: rankTransportOptions

prompt: |
  Choose which searches to run — do not
  always call both flights and trains.
Abbreviated from the sample. The full manifest includes input and output schemas; the model alias is configured by the application.

03 / Loomspan Console

See what happened.
Then follow the evidence.

A local companion to the framework. The Console connects to a selected Loomspan application and brings its runtime, skill catalog, and execution evidence into one place.

Console documentation
INSPECTION WORKFLOWLOCAL CONSOLE
  1. 01
    Orient to the application

    Runtime status, compatibility, and registered skills.

  2. 02
    Find the execution

    Active work, recent activity, and finalized traces.

  3. 03
    Follow the task hierarchy

    Plans, nested frames, and individual trace records.

  4. 04
    Inspect the underlying content

    Selected model and tool content, failures, and usage.

SEPARATE FROM YOUR APPLICATION

One native executable.

Written in Go with an embedded React interface. Node.js is needed to build the Console, not to run it.

THE SAME EVIDENCE, THROUGH MCP

A coding agent can investigate, too.

The Console exposes inspection tools through MCP. A companion agent skill guides runtime investigation, so an agent can follow observed executions while helping you develop and debug.

Explore the agent skills

04 / Getting started

Start in a Spring Boot application.

The repository contains the starter, a sample application, and the Console. Loomspan is under active development at 0.1.0-SNAPSHOT.

pom.xmlDEPENDENCY 
<dependency>
  <groupId>ai.loomspan</groupId>
  <artifactId>loomspan-spring-boot-starter</artifactId>
  <version>0.1.0-SNAPSHOT</version>
</dependency>
Java 21+ · Maven 3.9+ · Spring Boot 4.1 · Spring AI 2
  1. 01
    Configure the starter

    Declare model connections and skill locations.

  2. 02
    Define a skill

    Use a YAML manifest or a Java method.

  3. 03
    Invoke it from your application

    Inject SkillTemplate and call the skill by name.

ABOUT THE PROJECT

Loomspan was developed through human-directed collaboration with AI. Its approach to agent-assisted development extends to the framework itself: explicit skills, inspectable execution, and tools for working from runtime evidence.

Source and documentation