# Agent Scientist

> A terminal-based scientific AI assistant from Boltzmann Labs for literature research, data analysis, and approved computational workflows.

Agent Scientist is a scientific AI assistant from Boltzmann Labs that runs in the terminal. It supports researching literature, analyzing data, and running approved computational workflows, with reusable skills and connections to scientific services. The project is open source under the MIT license and installs from its GitHub repository.

## What It Is

Agent Scientist (launched with `agent-sci` or `sci`) is a command-line chat agent for life-science and related research tasks. Users chat with a configured model provider, and the agent can draw on skills, which are instructions for bioinformatics, chemistry, statistics, literature research, and visualization. Skills are instructions, not installed software, and the README notes they are not proof of scientific validity.

## Services, Skills and Profiles

Scientific tools can be connected through an API-key-gated Boltzmann connector, optional ToolUniverse and PubMed options, and other MCP-compatible services added with slash commands. The `/boltzmann_tools` command browses a catalog of scientific tools grouped by domain and module. Catalog visibility does not grant execution access.

Research profiles keep separate conversations, permissions, settings, and workspaces. A profile can combine selected skills, tools, service credentials, and optional local files or reference URLs. Users switch between profiles with `/activate`.

## Setup Path

Installation is done from a cloned repository using `install-sci.sh` on Ubuntu or `install-sci.ps1` on Windows. The installer prepares a managed runtime, locked dependencies, launchers, and bundled skills, then opens provider setup. `sci setup` and `sci model` configure model providers, including compatible custom endpoints. `sci doctor` checks the installation, and `sci update` updates a Git checkout.

## Safety Notes

Local execution is not an operating-system sandbox. Model weights, datasets, Docker images, and optional tool dependencies are not included. Computational outputs need scientific review, since successful execution does not establish experimental or clinical validity.

## Features
- Terminal-based scientific AI assistant
- Boltzmann scientific tool catalog (API-key gated)
- Optional ToolUniverse and PubMed MCP services
- Add other MCP services with /Add_mcp
- Reusable skills for bioinformatics, chemistry, statistics, literature research, and visualization
- Research profiles with separate conversations, permissions, settings, and workspaces
- Flexible model providers and compatible custom endpoints
- Persistent conversations with inspectable tool activity
- Resume saved sessions
- Capability inspection via /capabilities
- Self-update via sci update
- Skill, tool, and service connection extensibility

## Integrations
Boltzmann, ToolUniverse, PubMed, MCP servers, Codex

## Platforms
WINDOWS, MACOS, LINUX, WEB, API, CLI

## Pricing
Open Source

## Links
- Website: https://github.com/boltzmannlabs/Agent-Scientist
- Documentation: https://github.com/boltzmannlabs/Agent-Scientist#documentation
- Repository: https://github.com/boltzmannlabs/Agent-Scientist
- EveryDev.ai: https://www.everydev.ai/tools/agent-scientist