OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
Agent Infrastructure

Libra

Libra is an open-source Python systems framework for resource-aware, disaggregated and asynchronous post-training of agentic language models. It coordinates training and rollout clusters, routes requests across heterogeneous vLLM workers and adapts resource allocation to workload pressure during training.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Libra — OpenViST project profile in Agent Infrastructure.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use Libra when an agentic post-training workload needs disaggregated asynchronous execution and resource-aware coordination across training and rollout capacity.

Who it is for

  • AI infrastructure teams running agentic reinforcement-learning post-training
  • Researchers working with distributed GPU or NPU clusters
  • Engineers integrating vLLM rollout workers with FSDP or Megatron-based training

Documented uses

  • Coordinate training, rollout and elastic hybrid resource pools
  • Route agent requests across heterogeneous vLLM workers
  • Replan resource allocation as queue pressure and runtime workload change
  • Run asynchronous GRPO post-training workflows
  • Track planner decisions, throughput, rewards and reconfiguration events
Before you choose

Requirements & limits

Libra is publicly available as an open-source repository under the MIT license. It is a self-hosted Linux/server framework; no hosted access, pricing or account terms are specified here.

  • Whether the workload uses agentic RL post-training rather than general-purpose inference
  • Need for separate training, rollout and elastic hybrid pools
  • Compatibility with vLLM rollout workers and FSDP or Megatron training adapters
  • Availability of GPU or NPU cluster resources and operators able to configure distributed training
  • Value of adaptive replanning, heterogeneous tensor-parallel rollout buckets and runtime observability

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Prepare Libra for a self-hosted agentic RL setup

Set up the repository components and configuration needed to begin configuring agentic RL post-training.

Before you start

  • Linux or another documented server environment
  • Python-based Libra repository checkout
  • GPU or NPU cluster resources
  • vLLM, FSDP or Megatron-compatible training components as applicable
  • Model, hardware and experiment configuration

Steps

  1. Review the repository layout and select the relevant training, engine, workflow, environment and configuration components.

  2. Prepare hardware, model and experiment configuration using the repository configuration files.

  3. Select a documented training entry point or launcher under the examples and scripts directories.

  4. Configure the intended agentic workflow, tools, prompts, graders and rewards from the environment and workflow components.

Expected outcome

A configured Libra project structure ready for the next documented training or rollout stage.

Guide limitations

  • These sources document the setup components and entry-point locations but do not provide enough detail here to specify commands, exact configuration values or a completed run.
Documentation 1 ↗
COMPARE OPTIONS

Compare options for Agent orchestration

These projects document a common task and share a project role. This is a source-based comparison, not an execution benchmark or a guarantee of interchangeability.

ProjectDocumented purposePlatformAccess and requirements
LibraCoordinate training, rollout and elastic hybrid resource poolsLinux, Self-hosted, ServerLibra is publicly available as an open-source repository under the MIT license. It is a self-hosted Linux/server framework; no hosted access, pricing or account terms are specified here.
Agentfm CoreRun a containerized agent in a Podman sandbox on a worker machineDesktop, Docker, Self-hostedThe README lists desktop installers for macOS and Linux; it says the Windows app is in progress. Its desktop quick start says the bundled node connects to the public mesh. The README notes that the v1 installers are unsigned and macOS may block the first launch.
PlanoRoute incoming requests to agents based on their declared roles.Self-hosted, ServerPlano is public source code under the Apache-2.0 license. The README describes hosted Plano models free of charge in the US-central region for a first-run developer experience; for production scale it points to running models locally or contacting the project for API keys.
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
3 October 2026
Original website title
NetX-lab/Libra · Official homepage title · retained in its source language
How we review sourcesSuggest a correction