Repolode
Submit
beita6969

beita6969/ScienceClaw

An AI-for-science agent system on OpenClaw whose Skills and typed Operators grow from replay-verified executions, gated by independent validation.

View on GitHubWebsite
Stars
1,018
Forks
109
Open issues
10
Last commit
6 days ago

What is it?

ScienceClaw is an agent system for scientific tasks that accumulates reusable improvements. Each task is solved as a typed, executable workflow whose ports carry type, shape, unit and provenance information. When a repair reproduces under a clean environment reset, it is turned into a linked Skill (a strategy patch) and Operator (a typed, executable capability), committed together.

An update persists only if source replay reproduces and uses it, and independent validation tasks strictly improve within budget. Updates are versioned snapshots that users promote manually and can roll back. The repository includes a scientific tool library (42 modules, including forecasting, segmentation, protein and molecular models), a discipline skill set covering 23 research areas, and a gateway plugin for OpenClaw. A companion benchmark, ScienceClaw-Eval, is hosted on Hugging Face.

Key features

  • 01Workflows are typed graphs with schema-checked ports and fingerprinted nodes, so edits rerun only affected steps.
  • 02Reproduced repairs are split into Skill patches and Operators and committed as one bundle.
  • 03Candidate updates pass a replay-and-validation gate before persisting.
  • 04Every update is a versioned snapshot that can be promoted or rolled back via CLI.
  • 05Includes a 42-module scientific tool library and discipline skills for 23 research areas.
  • 06Exposes agent tools through an OpenClaw gateway plugin and a line-delimited JSON RPC interface.

Use cases

  • Running multi-step scientific analyses such as forecasting, segmentation or literature retrieval with verifiable outputs.
  • Building a reusable library of domain-specific workflows that improves as tasks are solved.
  • Benchmarking continual self-evolution of AI agents on scientific tasks.
  • Integrating scientific tool execution into an OpenClaw-based assistant.

In collections

Open-source AI agent frameworks & appsMCP servers

Topics

#ai-agent#llm#scientific-research#research-tools#self-evolving#mcp#bioinformatics#literature-review#ai#meta-analysis

FAQ

What is ScienceClaw?
An AI-for-science agent system on OpenClaw whose Skills and typed Operators grow from replay-verified executions, gated by independent validation. ScienceClaw is an agent system for scientific tasks that accumulates reusable improvements. Each task is solved as a typed, executable workflow whose ports carry type, shape, unit and provenance information. When a repair reproduces under a clean environment reset, it is turned into a linked Skill (a strategy patch) and Operator (a typed, executable capability), committed together.
Is ScienceClaw free to use? What license does it have?
ScienceClaw is open source under the MIT license. Check the license terms before using it in commercial products.
Is ScienceClaw still maintained?
Actively maintained. The last commit to ScienceClaw was on Oct 16, 2026.
What can I use ScienceClaw for?
Running multi-step scientific analyses such as forecasting, segmentation or literature retrieval with verifiable outputs. Building a reusable library of domain-specific workflows that improves as tasks are solved. Benchmarking continual self-evolution of AI agents on scientific tasks. Integrating scientific tool execution into an OpenClaw-based assistant.
Where can I get ScienceClaw?
The source code and installation instructions are on GitHub: https://github.com/beita6969/ScienceClaw

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects