Agentero is an MIT-licensed, TypeScript-built desktop research workspace for researchers who work alongside AI agents on PDF reading, cited question answering, vault-style notes, and writing, keeping the literature library and its structured context on the local machine.

What it is

Agentero is a unified research workspace for humans and agents, distributed as a desktop application for macOS, Windows, and Linux with a companion command-line tool. It combines a PDF reader with outline navigation, highlighting, and full-text search; a WYSIWYG Markdown editor that supports [[wikilink]] syntax, / shortcuts, and live LaTeX rendering; and an agent layer that connects to locally running agents over the Agent Client Protocol (ACP) and exposes the library to external clients as a Model Context Protocol (MCP) server. The project is written in TypeScript, carries the MIT licence, and has gathered 1,065 stars and 82 forks on GitHub.

The problem it addresses is the fragmentation of the modern research toolchain. Papers live in one application, notes in another, and the reasoning done with a model lives in assorted web chat windows, so a researcher who wants an agent to work through a specific paper alongside earlier notes has to copy and paste by hand. Agentero also targets the mismatch between human-readable PDFs and model-readable content: two-column layouts, tables spanning pages, formulas, and cross-references are costly and error-prone for models to parse, whereas structured Markdown and LaTeX are not. The README is explicit that Agentero is not intended to replace Zotero; it explores a different arrangement in which humans set research questions and judge academic value while agents handle batch retrieval, metadata capture, figure and formula extraction, citation tracing, and basic summarisation.

Key capabilities

• Structured extraction of mathematical formulas, algorithm pseudocode, and tables from papers into Markdown and LaTeX, so agent context is accurate.
• Agent connectivity over ACP to local agents such as Claude Code, OpenCode, Codex, or local models, with context kept on the machine and no vendor lock-in.
• A built-in MCP server that lets ChatGPT Web and Claude Desktop query and use the local literature library as context.
• A headless CLI for scripted batch import of papers, metadata extraction, and note writing.
• Discovery through built-in Cool Papers and ModelScope integration, academic RSS subscriptions, and daily arXiv recommendations based on the existing library.
• Automatic parsing of reference lists with one-click batch fetching, plus publisher metadata retrieval and citation tracking.
• Zotero library import preserving tags, notes, and attachments, compatibility with the Zotero browser extension, and BibTeX/BibLaTeX export.

Who uses it and how

• Individual researchers who read PDFs on a laptop while running a local agent for questioning and derivation, with highlights on the left and answers on the right.
• Users with an existing Zotero archive who want to migrate the whole library, keep their tags and notes, and continue saving papers through the Zotero browser extension.
• Researchers who want the same library on several devices and configure an S3-compatible storage service for sync.
• Users who run the library on a remote server and reach it through an SSH tunnel.
• Automation-minded users who script bulk imports and metadata capture through the CLI instead of saving papers one at a time in a browser.

Getting started

Prebuilt installers for macOS, Windows, and Linux are available from agentero.app and GitHub Releases; on macOS the application installs with brew tap poco-ai/agentero followed by brew install --cask agentero, and the CLI with brew install agentero.

How it compares

No list of paid products is given in the facts, so the useful reference point is Zotero, which Agentero explicitly does not set out to replace: it instead adds an agent-facing layer and keeps compatibility through library import, the browser extension, and BibTeX export. It differs from that established toolchain in being MIT-licensed, storing everything as ordinary PDF and Markdown files in plain local folders rather than a private database or hashed directory, and in treating model context as a first-class output.

When to use it — and when not to

Choose Agentero if you already work with local agents or external AI clients and want papers, notes, and interaction records kept in one transparent local store. Be prepared to operate the surrounding pieces yourself: a local agent or agent runtime, an S3-compatible service if you want multi-device sync, and optional BYOK translation keys; Linux requires Ubuntu 22.04 or newer with webkit2gtk 4.1. It is a poor fit for anyone who wants a hosted web service with no local setup, and the project is early-stage, with 38 open issues at the time of writing.

brew tap poco-ai/agentero
brew install --cask agentero

brew tap poco-ai/agentero
brew install agentero

# 查看常用命令与参数
agentero --help