OpenSpace: HKU’s Open‑Source AI Agent Engine Cuts Tokens by 46% and Boosts ROI 4.2×

One‑sentence overview

OpenSpace is a self‑evolving AI agent engine that enables agents to learn from real‑world experience, adapt, evolve, and share skills to create collective intelligence.

Core value

Less Token

: reuses successful patterns to avoid reasoning from scratch each time. Self‑evolution Skills: automatic skill repair, improvement, and learning. Agent Experience Sharing: one agent learns, all agents benefit.

Current AI agent pain points

Three core problems

1. Massive token waste

Each inference starts from zero, preventing reuse of successful task patterns.

Token consumption repeats the same work.

2. Repeated expensive failures

Solutions cannot be shared between agents.

The same costly exploration and errors are repeated.

3. Unreliable skills

When tools or APIs evolve, skills silently degrade.

Community‑contributed skills lack strict quality standards.

OpenSpace solution

Self‑evolution capability

AUTO‑FIX

: instantly repairs damaged skills. AUTO‑IMPROVE: converts successful patterns into better skill versions. AUTO‑LEARN: captures successful workflows from real usage. Quality Monitoring: tracks skill performance, error rates, and success rates across tasks.

Continuous evolution turns every failure into improvement and every success into optimization.

Collective agent intelligence

Shared Evolution

: an improvement in one agent upgrades all agents. Network Effect: more agents generate richer data, accelerating evolution for each agent. Easy Sharing: a single command uploads or downloads evolved skills. Access Control: each skill can be public, private, or team‑only.

On 50 professional tasks across six industries (GDPVal benchmark) using the same backbone LLM (Qwen 3.5‑Plus), OpenSpace agents achieved a 4.2× return while reducing token usage by 46%.

Quick start

Installation

git clone https://github.com/HKUDS/OpenSpace.git && cd OpenSpace
pip install -e .
openspace-mcp --help  # verify installation

Method A: Integrate into your agent

Applicable to any agent that supports a SKILL.md file (e.g., Claude Code, Codex, OpenClaw, nanobot).

Add OpenSpace to the agent’s MCP configuration:

{
  "mcpServers": {
    "openspace": {
      "command": "openspace-mcp",
      "toolTimeout": 600,
      "env": {
        "OPENSPACE_HOST_SKILL_DIRS": "/path/to/your/agent/skills",
        "OPENSPACE_WORKSPACE": "/path/to/OpenSpace",
        "OPENSPACE_API_KEY": "sk-xxx (optional for cloud services)"
      }
    }
  }
}

Copy the provided skills into your agent’s skill directory:

cp -r OpenSpace/openspace/host_skills/delegate-task/ /path/to/your/agent/skills/
cp -r OpenSpace/openspace/host_skills/skill-discovery/ /path/to/your/agent/skills/

These two skills teach the agent when and how to use OpenSpace.

Method B: Use as an AI colleague directly

# interactive mode
openspace
# execute a task
openspace --model "anthropic/claude-sonnet-4-5" --query "Create a monitoring dashboard for my Docker container"

Local dashboard

View skill evolution – browse skills, track lineages, compare differences.

# terminal 1: start backend API
openspace-dashboard --port 7788
# terminal 2: start frontend dev server
cd frontend
npm install
npm run dev

Architecture design

Self‑evolution engine

Skills are living entities that can be automatically selected, applied, monitored, analyzed, and evolved.

Three evolution modes FIX: in‑place repair of broken or outdated commands, creating a new version of the same skill. DERIVED: create enhanced or specialized versions from a parent skill; the new skill coexists with its parent. CAPTURED: extract reusable patterns from successful executions; the new skill has no parent.

Three independent triggers Post‑execution analysis: runs after each task, analyses the full log and suggests evolution. Tool degradation: when a tool’s success rate drops, batch‑evolve all dependent skills. Metric monitoring: periodically scans skill health metrics and evolves under‑performing ones.

Showcase case

The case starts from zero code, evolves 60+ skills, and builds a fully functional real‑time dashboard.

🌱 Seed : analyze the open‑source WorldMonitor and extract reference patterns – 6 initial skills.

🏗️ Scaffold : generate project structure, Vite config, TypeScript setup – +8 skills.

🎨 Build : create 20+ panels, data services, API routes, grid layout – +25 skills.

🔧 Fix : automatically fix broken TypeScript, API mismatches, CSS conflicts – +12 FIX evolutions.

🧬 Evolve : derive enhanced patterns and merge complementary skills – +15 DERIVED skills.

📦 Capture : extract reusable patterns from successful runs – +8 CAPTURED skills.

Reference resources

GitHub repository: https://github.com/HKUDS/OpenSpace

Official site: https://open-space.cloud

Chinese README: README_CN.md in the repository

GDPVal benchmark: https://huggingface.co/datasets/openai/gdpval