EvoScientist/EvoScientist

EvoScientist – Self‑Evolving AI Research Assistant

What it is – EvoScientist is a Python‑based framework that lets a team of LLM‑driven agents act as an autonomous “AI scientist”. The agents collaborate to plan, research, write code, debug, analyse data and draft papers, while continuously improving their own memory and skills.

Core ideas

• Human‑on‑the‑loop – the system runs on its own but a human can intervene, approve actions, or schedule recurring research runs.
• Self‑evolving memory – observations are auto‑distilled into a knowledge‑graph that grows across sessions; recurring patterns are turned into reusable AutoSkills.
• Multi‑provider, multi‑channel – works with dozens of LLM providers (Anthropic, OpenAI, Google, MiniMax, NVIDIA, etc.) and can be accessed via a desktop Web UI, a CLI/TUI, or messaging platforms such as Telegram, Slack, WeChat, etc.
• Full research pipeline – intake → planning → execution → evaluation → writing → verification, all orchestrated by six sub‑agents (plan, research, code, debug, analyse, write).

Key features

Feature	What it does
Multi‑Agent Team	Six specialised agents run in parallel, sharing observations via a LangGraph‑backed workflow.
Self‑Evolving Memory	Observations are linked into a knowledge graph; the graph is auto‑distilled each turn.
AutoSkills	Periodically extracts repeatable patterns into installable skill modules (/autoskills).
Multi‑Provider Model Switching	One config file lets you swap between Anthropic, OpenAI, Google, MiniMax, NVIDIA, etc.
Multi‑Channel Access	Same session reachable from CLI, Web UI, Telegram, Slack, Feishu, WeChat, …
Desktop Web UI	Browser‑based workspace that mirrors the CLI session (--ui webui).
Scheduled / Cron‑style Tasks	Define recurring research jobs that run unattended and report back.
Adaptive Tool Selection	Only the most relevant tools are exposed each turn, reducing clutter.
Context Editing	System prompt is rewritten on‑the‑fly based on conversation state.
MCP & Skill Plug‑ins	Load external “Micro‑Control‑Program” servers or install community‑made skills from GitHub.

Typical use cases

• Literature review & hypothesis generation – the research sub‑agent scans papers, extracts insights, and proposes new ideas.
• Prototype code & debugging – the code and debug agents write, run, and fix scripts automatically.
• Data analysis pipelines – the analyse agent loads datasets, runs statistical or ML analyses, and produces visualisations.
• Manuscript drafting – the write agent composes sections, formats citations, and iteratively refines a paper.
• Long‑running projects – schedule a weekly “run” that lets the system explore a topic, generate code, and email a summary.

Getting started

1. Install via PyPI: pip install EvoScientist (current version 0.3.3).
2. Run the CLI: evo sci start --ui webui to launch the browser UI, or simply evo sci start for the terminal interface.
3. Configure your LLM providers in ~/.evo_scientist/config.yaml (API keys for OpenAI, Anthropic, etc.).
4. Use slash‑commands such as /plan <topic>, /schedule “every Monday”, or /expert <skill> to steer the agent.
5. Review generated AutoSkills with /autoskills and install community packs from the companion EvoSkills repository.

Documentation & community

• Website – https://EvoScientist.github.io/ (demos, tutorials, API reference).
• Technical report – arXiv:2603.08127 (details the architecture and benchmark results).
• Benchmarks – consistently ranked #1 on DeepResearch Bench, AstaBench, and ResearchClawBench (2025‑2026).
• License – Apache 2.0 (permissive, commercial‑friendly).

Why it matters EvoScientist bundles recent advances in LLM‑driven tool use, LangGraph orchestration, and memory‑graph distillation into a ready‑to‑run research assistant. By automating repetitive scientific tasks while still allowing human oversight, it aims to accelerate discovery and lower the barrier for researchers to experiment with AI‑augmented science.