HKUDS/OpenSpace: "OpenSpace: Make Your Agents: Smarter, Low-Cost, Self-Evolving" -- Community: https://open-space.cloud/

"OpenSpace: Make Your Agents: Smarter, Low-Cost, Self-Evolving" -- Community: https://open-space.cloud/ - HKUDS/OpenSpace

Summary

OpenSpace is a Python package (MIT licensed, Python 3.12+) that acts as a "Skill Management Layer" for AI agents, plugging in via MCP to retrieve, evaluate, share, and evolve reusable skills. It targets the situation where an agent's skill library grows large but the agent cannot tell which skills actually work. The project ships a CLI, MCP server, local dashboard, and TUI, and lists Claude Code, Codex, OpenClaw, Hermès, and nanobot as supported hosts.

What it is useful for

OpenSpace addresses four practical gaps when working with skill-driven agents:

• Finding the right skill at scale: as skill libraries grow into the hundreds, the README argues the right entry becomes harder to surface. OpenSpace indexes local skills and supports lexical plus semantic-style ranking, with an embedding cache that speeds up repeat searches.
• Quality evidence instead of trust-by-description: every run records whether a skill was selected, applied, completed, or fell back to a fallback. Package and skill detail views surface these quality summaries alongside lineage and history.
• Controlled evolution: candidate improvements are marked provisional, validated against real outcomes, and only promoted once they earn trust, with version history kept for every change.
• Local-first sharing: cloud skills are browsed as packages, then explicitly imported into a local skill folder before reuse, so execution and data stay on the host.

How engineers can use it

The README documents two paths after git clone and pip install -e ..

Path A — plug it into your agent via MCP. Register the server in your host agent's MCP config:

{
  "mcpServers": {
    "openspace": {
      "command": "openspace-mcp",
      "toolTimeout": 600,
      "env": {
        "OPENSPACE_HOST_SKILL_DIRS": "/path/to/your/agent/skills",
        "OPENSPACE_WORKSPACE": "/path/to/OpenSpace"
      }
    }
  }
}

Then copy two host skills into your agent's skills 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/

For remote hosts, openspace-mcp also runs over SSE or streamable HTTP instead of stdio. Credentials are auto-detected from nanobot and OpenClaw; other hosts should set OPENSPACE_LLM_API_KEY and OPENSPACE_MODEL, or populate openspace/.env.

Path B — use the CLI directly:

openspace --model "anthropic/claude-sonnet-4-5" --query "Create a monitoring dashboard for my Docker containers"

Custom skills live in .openspace/skills/<skill-name>/SKILL.md. Cloud features (browsing community packages, sharing) require running openspace-cloud-auth bootstrap-agent-key to provision an API key; without it, local execution, evolution, and search still work.

Documented limitations and caveats: the default clone is large because the assets/ folder adds roughly 50 MB, so the README recommends git clone --filter=blob:none --sparse and skipping assets/. litellm is pinned below 1.82.7 to avoid a known supply-chain issue, and Windows users should use the fixed stdio and gateway PID handling shipped in recent updates. Per-agent configuration beyond the two paths above is in openspace/host_skills/README.md.