What is OpenSpace?
OpenSpace, from the University of Hong Kong's Data Intelligence Lab (HKUDS), is a Python layer that sits beside Claude Code, Codex, OpenClaw, nanobot or any MCP-capable agent and manages its SKILL.md skills. It records whether each skill was selected, applied, completed the task or fell back, uses that evidence to fix, derive or capture skills automatically, and offers a cloud hub at open-space.cloud for sharing them.
Version 2.0.0, released on 17 July 2026, added quality summaries, package browsing, task-trace uploads, a dashboard and a terminal UI. New or evolved skills start as provisional and are promoted to trusted after independent successful use. The team reports, on its own Terminal-Bench 2.1 runs with the same Hy3 model, 65.2% with a fresh skill library and 78.7% once it had evolved.
The repository was created on 24 March 2026, is MIT-licensed and has about 7,800 GitHub stars. The last commit was on 12 August 2026, and it is installed from a clone rather than a published package.
Is OpenSpace worth it?
OpenSpace is built on a sound observation: as a skill library grows past a few dozen entries, agents pick the wrong one, keep using ones that fail, and nobody can tell good skills from plausible-looking ones. Recording outcomes per skill, and preferring skills with a record of completing tasks, is a better signal than a description, and the provisional-to-trusted lifecycle means an evolved skill has to succeed again, independently, before it is treated as reliable. Skills pass a safety scan for prompt-injection and credential-exfiltration patterns before loading.
The headline benchmark needs its context. The jump from 65.2% to 78.7% on Terminal-Bench 2.1 is the team's own result with one model, the warm run benefiting from a library evolved on related tasks; it shows the mechanism can help, not that it will on your workload. Integration is via an MCP server that wraps OpenSpace's own execution agent, so long delegated tasks need a host timeout of at least 600 seconds.
The defaults are what to check first. Evolution evidence collection is on by default, and the README's configuration notes (currently commented out) describe autonomous mode, in which validated fixes, derived and captured skills commit without a person, as the default; audit_only and fix_only modes exist and suit most teams better. The bundled telemetry module, inherited from mcp-use, sends anonymised events to PostHog and Scarf unless MCP_USE_ANONYMIZED_TELEMETRY=false is set. Cloud features are opt-in and need an agent key tied to an email address.
Maintenance is the other concern. After a busy spring, the last commit landed on 12 August 2026 and v2.0.0 remains the latest release, with about 54 open issues and pull requests. There is no PyPI package, and a default clone drags in about 50 MB of images. Compared with Hermes Agent, which also writes its own skills but as part of one maintained product, OpenSpace is a layer you bolt on, with more evidence tracking and less certainty that it will be updated.
Where it earns its place
A platform team maintaining 300 internal skills across Claude Code and Codex
Find out which skills actually help and retire the rest
Run OpenSpace as an MCP server in audit_only mode so it records selection and outcome per skill without changing any, then review quality summaries in the dashboard. Free; cost is the model calls its post-task analysis makes.
An AI engineer at a 10-person automation agency
Turn a workflow that worked for one client into a reusable, tracked skill
Let CAPTURED evolution save the validated sub-workflow as a provisional skill, promote it after a second success, and share it privately through a cloud group. Set MCP_USE_ANONYMIZED_TELEMETRY=false first.
A researcher studying self-improving agents
Reproduce the effect of an evolving skill library on a coding benchmark
Clone the repository, run cold and warm passes with the same model, and inspect lineage in the dashboard. The published 65.2% to 78.7% figure is the team's own, so treat it as a hypothesis to test.
Before you commit
- Setup
- git clone and pip install -e . on Python 3.12+, add an openspace MCP server to the host config with a 600-second tool timeout, and copy two host skills into the agent's skill folder. The dashboard needs Node.js 20+.
- Integrations
- Any MCP host over stdio, SSE or streamable HTTP; helpers for OpenClaw and nanobot, manual setup for Claude Code, Codex and Cursor; WhatsApp and Feishu gateway adapters; models through LiteLLM including Ollama.
- Data and privacy
- Skills, evidence and lineage live in a local SQLite store. Anonymised PostHog and Scarf telemetry is on unless MCP_USE_ANONYMIZED_TELEMETRY=false; cloud upload is opt-in, needs an email-linked agent key and accepts only trusted skills. MIT licence.
- Cost at scale
- No paid tier published. Each task adds post-execution analysis and possible evolution calls on your own model key, on top of the work itself.
- Limits
- Latest release v2.0.0 (July 2026) and no commit since 12 August; about 54 open issues and pull requests; no PyPI package; large clone because of about 50 MB of assets.
What’s good
- Scores skills on recorded task outcomes, not descriptions
- Provisional skills must succeed again before trusted
- Works with any MCP host that reads SKILL.md
- Safety scan blocks injection-pattern skills on load
What’s not
- Autonomous skill rewriting described as the default
- PostHog and Scarf telemetry on unless disabled
- No commit since 12 August 2026
- Benchmark gains are self-reported on one model
Best for
Teams running agents against a large, growing skill library who want evidence about which skills work and will set evolution to audit-only first.
OpenSpace alternatives
OpenSpace: common questions
- Is OpenSpace worth it?
- OpenSpace asks the right question about agent skills, which ones actually work, but it lets agents rewrite their own skills by default and has gone quiet since August. We rate it 3.5 out of 5.
- How much does OpenSpace cost?
- Free and open source (MIT). The open-space.cloud hub is optional and no paid plan is published. You pay your model provider.
- What are the downsides of OpenSpace?
- Autonomous skill rewriting described as the default; postHog and Scarf telemetry on unless disabled; no commit since 12 August 2026; benchmark gains are self-reported on one model.
- Who is OpenSpace best for?
- Teams running agents against a large, growing skill library who want evidence about which skills work and will set evolution to audit-only first.
- How much does OpenSpace cost for a team or in production?
- No paid tier published. Each task adds post-execution analysis and possible evolution calls on your own model key, on top of the work itself.
- Is OpenSpace safe for company data?
- Skills, evidence and lineage live in a local SQLite store. Anonymised PostHog and Scarf telemetry is on unless MCP_USE_ANONYMIZED_TELEMETRY=false; cloud upload is opt-in, needs an email-linked agent key and accepts only trusted skills. MIT licence.
- What are the best OpenSpace alternatives?
- The closest alternatives we have reviewed are Hermes Agent, claude-mem and Everything Claude Code (ECC).
Answers checked Oct 31, 2026.