EvoScientist is an Apache-2.0 Python research platform that runs a self-evolving team of AI scientists — six coordinated sub-agents with persistent memory and installable skills — for researchers and research engineering teams who want autonomous exploration, code execution, and write-up driven from one CLI, desktop WebUI, or chat channel.

What it is

EvoScientist sits in the AI & Machine Learning / AI Development Platforms category and is tagged ai-agent, ai4science, multi-agent-system, and vibe-research. Its core is a multi-agent team of six sub-agents — plan, research, code, debug, analyze, and write — working under a unified control surface. The memory layer is self-evolving: it is auto-distilled each turn and self-links into a knowledge graph that grows across sessions. AutoSkills distills recurring patterns out of that memory into reusable skills on a schedule and proposes them for review through /autoskills. Research follows a fixed scientific workflow of intake, plan, execute, evaluate, write, and verify. The project describes itself as opinionated and ready to use out of the box, and it adopts a human-on-the-loop paradigm rather than human-in-the-loop, positioning the agent as a research buddy that co-evolves with human researchers and internalizes scholarly taste and scientific judgment.

The concrete problem it addresses is the friction of driving a research assistant turn by turn while stitching together separate chat tools, notebooks, and scripts. EvoScientist replaces that human-in-the-loop arrangement with one agent session reachable from the CLI, a desktop WebUI, or chat channels, supported by adaptive per-turn tool selection that keeps only relevant tools visible and context editing that rewrites the system prompt as the conversation changes. Scheduled tasks run recurring research on a cron-style schedule without supervision and report back. Reported results include first place at submission time on DeepResearch Bench, DeepResearch Bench II, AstaBench Code & Execution, and AstaBench Data Analysis, second place overall on ResearchClawBench Agent Mode (first among GPT-5.4-based agents), a Best Paper & Appraisal Award, and a technical report on arXiv 2603.08127.

Key capabilities

• Six sub-agents cover the research lifecycle: plan, research, code, debug, analyze, and write.
• Self-evolving memory is auto-distilled each turn and self-links into a knowledge graph that persists across sessions.
• AutoSkills distills recurring patterns from memory into reusable skills on a schedule and proposes them for review via /autoskills.
• Multi-provider support covers Anthropic, OpenAI, Google, MiniMax, and NVIDIA through one config to switch.
• Multi-channel delivery treats the CLI as the hub, with Telegram, Slack, Feishu, WeChat, and more sharing one agent session.
• A desktop WebUI workspace-panel app launches from a terminal with --ui webui.
• MCP servers can be plugged in and skills installed from GitHub on the fly, with the companion EvoSkills repository covering the end-to-end research lifecycle.

Who uses it and how

• Researchers running a full intake, plan, execute, evaluate, write, and verify pipeline on a single question.
• Teams that need recurring research executed unattended and reported back on a cron-style schedule.
• Groups already working inside Telegram, Slack, Feishu, or WeChat who want one agent session across all of them with the CLI as hub.
• Desktop users who prefer the workspace-panel WebUI launched with --ui webui over a terminal.
• Agent builders who plug in MCP servers, install GitHub skills, or use the EvoSkills set, which is also compatible with other CLI coding agents.

Getting started

Launch runs from the command line, with --ui webui starting the desktop WebUI and one config selecting the provider among Anthropic, OpenAI, Google, MiniMax, and NVIDIA. The supplied excerpt names no package name, Docker image, or compose file, so installation steps live on https://EvoScientist.ai/ and in the repository README.

How it compares

No list of paid products this project replaces is provided in the facts, and no directly comparable tools are named — the only adjacent reference is compatibility with other CLI coding agents, which EvoSkills supports rather than competes with. On the available facts, EvoScientist stands alone in this registry as an AI4Science multi-agent platform.

When to use it — and when not

A self-hoster must operate provider API credentials for at least one of the supported vendors, the memory and knowledge-graph store, and any chat or MCP integrations they enable, which is real operational surface for a single-researcher setup. Teams that require strict human-in-the-loop approval at every step, or that will not review AutoSkills proposals, should look elsewhere because the design assumes human-on-the-loop oversight. The excerpt also ships no installation command, package name, or container image, so a prospective adopter must check the homepage and repository first.

</td>

</td>

</td>