Evoscientist
🔬 Harness Vibe Research with Self-evolving AI Scientists
Last verified:
What is Evoscientist?
EvoScientist is a self-evolving multi-agent AI scientist framework designed for end-to-end scientific discovery. It harnesses
Evoscientist pricing
Pricing model: Freemium
EvoScientist is free and open-source under Apache 2.0 license. No paid tiers mentioned. Requires users to provide their own LLM provider API keys (Anthropic, OpenAI, Google, NVIDIA, SiliconFlow, OpenRouter, Volcengine, DashScope, or Ollama). Optional OAuth sign-in available for Claude Code and Codex CLI subscribers which may not require API key. Channel integrations require extra dependencies that can be installed selectively (telegram, discord, slack, wechat, qq, feishu) or all-at-once with [all-channels].
Evoscientist pros
- Ranked #1 on DeepResearch Bench at submission time
- Ranked #1 on DeepResearch Bench II at submission time
- Ranked #1 on AstaBench Code & Execution
- Ranked #1 on AstaBench Data Analysis
- Best Paper Award at ICAIS 2025 AI Scientist Track with 6/6 papers accepted
- 6 specialized sub-agents working in concert under LangGraph
- Persistent memory preserves context across sessions
- Supports 9 LLM providers with one config to switch
- 10 messaging channel integrations (Telegram, Discord, Slack, WeChat, etc.)
- 13 research-lifecycle EvoSkills covering ideation to publication
- Sandboxed code execution with 300s timeout and auto-recovery
- Human approval for high-risk tool calls before execution
- Voice input support via faster-whisper (zh/en/auto)
- 14 slash commands for session management
- MCP server integration for extending toolsets
- Async sub-agents for true parallel multi-agent execution
- Adaptive context with per-turn tool filtering
- Auto-detect model names or specify full IDs directly
- Docker support with sandboxed shell access
- Apache 2.0 open-source license
Evoscientist cons
- Requires Python 3.11+ but less than 3.14
- iMessage channel not usable from Docker container
- Shell commands require human approval by default
- Code execution has 300s timeout limit
- Output limits in sandboxed workspace
- STT and OAuth not bundled, must install on demand
- TinyTeX/LaTeX not bundled, must install separately for paper-writing
- Container runs as non-root user requiring directory permission setup
- Web app with workspace UI not yet released (roadmap item)
- Scheduled tasks for core system not yet implemented
Frequently asked questions about Evoscientist
What is EvoScientist?
EvoScientist is a self-evolving multi-agent AI scientist framework for end-to-end scientific discovery. It harnesses vibe research by enabling AI scientists that autonomously explore, generate insights, and iteratively improve through persistent memory and self-evolution. The system uses 6 specialized sub-agents (plan, research, code, debug, analyze, write) working under a shared LangGraph state machine.
How do I install EvoScientist?
The quickest way is to run 'uv tool install EvoScientist'. Requirements are Python 3.11+ (less than 3.14). After installation, run 'EvoSci onboard' to configure your LLM provider, API keys, model selection, and workspace mode via an interactive wizard. Alternatively, you can use Docker with the pre-built image from ghcr.io/evoscientist/evoscientist:latest.
Which LLM providers does EvoScientist support?
EvoScientist supports 9 LLM providers: Anthropic (claude-opus-4-6, claude-sonnet-4-6, claude-haiku-4-5), OpenAI (gpt-4o, o3-mini, o1), Google (gemini-2.5-pro, gemini-2.5-flash, gemini-2.0-flash), NVIDIA (deepseek-r1, llama-3.3-70bn, emotron-ultra), SiliconFlow (deepseek-v3, qwen-plus, glm-4), OpenRouter (any model), Volcengine (doubao-pro, doubao-lite), DashScope (qwen-max, qwen-turbo), and Ollama/Custom for local or self-hosted models.
What messaging channels are supported?
EvoScientist supports 10 messaging channels: iMessage, Telegram, Discord, Slack, WeChat, DingTalk, Feishu, Email, QQ, and Signal. The CLI serves as the hub, and all channels share the same agent session. Setup is done via 'EvoSci channel setup <channel>' and multiple channels can run concurrently.
What are EvoSkills?
EvoSkills is the official skill repository containing 13 research-lifecycle skills covering the full pipeline from ideation to publication. Skills can be installed all at once with a single command and are compatible with Claude Code, Cursor, and other AI coding agents. Skills extend EvoScientist's functionality for specific research tasks.
How does persistent memory work?
EvoScientist's persistent memory preserves context, preferences, and experimental findings across sessions. The system internalizes scholarly taste and builds on prior work. Memory includes ideation memory (summarizing feasible research directions and recording unsuccessful ones) and experimentation memory (capturing effective data processing and model training strategies).
What is the scientific workflow in EvoScientist?
EvoScientist follows a 6-phase scientific workflow: Intake → Plan → Execute → Evaluate → Write → Verify. It uses baseline-first design with one-variable iteration for scientific rigor. The workflow is orchestrated by the multi-agent team with every step verified and every result verified.
How do I handle code execution approvals?
By default, shell commands require human approval before running. You can skip approval prompts by using --auto-approve flag, setting auto_approve true in config via 'EvoSci config set auto_approve true', or allowing specific command prefixes like 'python,pip,pytest,ruff,git'. During a session, reply 3 (Approve all) at any approval prompt to auto-approve for the rest of that session.
What benchmarks has EvoScientist achieved?
EvoScientist ranked #1 on DeepResearch Bench (Apr 2026), #1 on DeepResearch Bench II (Apr 2026), #1 on AstaBench Code & Execution (Mar 2026), and #1 on AstaBench Data Analysis (Mar 2026). It also won Best Paper Award and AI Reviewer's Appraisal Award at ICAIS 2025 AI Scientist Track with 6/6 papers accepted, including an AI-generated best paper.
What is the license and how can I contribute?
EvoScientist is licensed under Apache License 2.0. The project welcomes contributions from developers, researchers, and AI coding agents. You can join the community via Discord (for real-time questions and collaboration) or WeChat (for Chinese-speaking research community). For inquiries, contact [email protected]. The GitHub repository has a Contributing Guidelines covering architecture, patterns, extension guides, and code standards.