HKUDS/OpenSpace

What it solves

AI agents often struggle to manage a growing library of skills, leading to poor retrieval of the right tool for a task, repeated mistakes, and a lack of of quality signals to distinguish between reliable and ineffective skills. OpenSpace provides a management layer that allows agents to retrieve, evaluate, and evolve their capabilities based on real-world outcomes.

How it works

OpenSpace acts as a skill management layer that plugs into agents via the Model Context Protocol (MCP). It implements a lifecycle for agent skills through four primary capabilities:

• Evaluation: It tracks every execution run to monitor whether a skill was successfully applied or replaced by a fallback, using real task outcomes as evidence of quality.
• Evolution: It allows skills to be updated (FIX, DERIVED, or CAPTURED) based on evidence, with version history and validation to ensure improvements are controlled.
• Local-First Hub: It uses a local-first architecture where skills are executed locally, but can be discovered and imported from a cloud-based package system.
• Agent Harness: It provides a runtime environment that captures quality records and manages tool permissions and sandboxing.

Who it’s for

Developers building AI agents (such as those using Claude Code, Codex, or OpenClaw) who need a scalable way to manage, share, and improve the skills their agents possess over time.

Highlights

• Evidence-Driven Quality: Skills are ranked and trusted based on actual task completion rates rather than just descriptions.
• Self-Evolution: Agents can improve their own skill library from a "cold run" to a "warm run" by learning from successful workflows.
• MCP Integration: Compatible with any host that supports the Model Context Protocol.
• Private Infrastructure: Supports private deployment to keep workflows and data under the user's own control.