What is Corbis (corbis.ai)

Corbis is research-first AI for finance, real estate, and economics. You ask in plain language; it searches a large, domain-specific index (hundreds of thousands of peer-reviewed papers, plus industry reports and market data in the product) and returns answers with citations you can open and check. The goal is evidence you can stand behind, not unattributed claims. For a journal-level snapshot of publications and what the corpus includes, see Research Insights.

The full platform adds chat, guided workflows, and exports (for example PDF, Word, LaTeX, and citation formats). MCP (Model Context Protocol) exposes the same underlying tools to external clients (Cursor, Claude Code, Codex, and other compatible agents). This repository is a literature starter kit for that path: it focuses on surveying literature, mapping a field, screening ideas, and keeping citations and search trails explicit in your repo.

Why This Exists

Most AI assistants are good at sounding confident and bad at doing literature work carefully. Corbis fixes the first part by giving them live research, data, and citation tools through MCP. This repo fixes the second part by packaging those tools into reusable research workflows.

Use it to:

• survey a literature fast
• find the closest papers before you pitch an idea
• brainstorm ideas and kill weak ones early
• export citations without formatting them by hand
• generate literature figures and keep a transparent search trail

It works with Codex, Claude Code, Cursor, and other MCP-compatible agents.

Claude Code gets the smoothest out-of-the-box slash-command experience. Codex, Cursor, and other MCP clients can use the same Corbis tools and the same workflow prompts from this repo.

The Workbench

These are the six workflows bundled with the kit:

Plus a paper-reader agent prompt for assistants that support repo-defined agents.

If your client does not support slash commands directly, use the same workflow names as prompt starters or follow the examples in SKILLS_USE_GUIDE.md.

Quick Setup

You need two things: an AI assistant with MCP support and a Corbis MCP API key.

1. Get a Corbis API key

Open the Corbis app, go to Settings > API Keys, and create a key.

Corbis MCP keys start with corbis_mcp_. Copy the key when it is created. It is shown once.

2. Connect your assistant

Add Corbis to ~/.codex/config.toml for global use, or .codex/config.toml for a project-local setup:

[mcp_servers.corbis]
url = "https://www.corbis.ai/api/mcp/universal"
bearer_token_env_var = "CORBIS_MCP_API_KEY"
startup_timeout_sec = 20
tool_timeout_sec = 120

Then export your key before starting Codex:

export CORBIS_MCP_API_KEY="corbis_mcp_..."
codex

Full guide: CORBIS_MCP_CODEX_GUIDE.md

export CORBIS_MCP_API_KEY="corbis_mcp_..."
git clone https://github.com/Agentic-Assets/corbis-literature-starter-kit.git my-project
cd my-project && claude

The committed .mcp.json uses an Authorization header and reads the key from the environment.

Full guide: CORBIS_MCP_CLAUDE_CODE_GUIDE.md

git clone https://github.com/Agentic-Assets/corbis-literature-starter-kit.git my-project

Then connect the URL https://www.corbis.ai/api/mcp/universal in Cursor and authenticate with OAuth or an Authorization header. See CORBIS_CURSOR_PLUGIN.md for setup.

Open the project after connecting the server. Cursor can use the same Corbis MCP tools and repo guidance.

Connect to this MCP endpoint:

https://www.corbis.ai/api/mcp/universal

Authenticate through OAuth when supported, or send Authorization: Bearer YOUR_KEY in a header. Keep keys out of URLs.

Architecture and client notes: CORBIS_MCP_GUIDE.md

No Python is required for search, review, idea screening, or citation workflows. Python is only needed for the figure-generation workflow.

First Prompts Worth Stealing

Map a field before you pretend to know it

/lit-review climate risk and commercial real estate pricing

Builds a paper set, clusters the literature into themes, and writes a synthesized review with citations.

Pressure-test an idea before it eats a month of your life

/idea Do bank branch closures reduce small business lending through relationship destruction?

Finds the closest papers, scores the idea across multiple dimensions, and returns a go, revise, or kill recommendation.

Generate ideas, but keep only the survivors

/brainstorm behavioral biases in household mortgage decisions

Creates a wider internal idea pool, rejects weak candidates, and ranks the survivors by novelty, importance, and executability.

Turn a messy literature into figures

/lit-landscape corporate governance and firm performance

Produces timelines, landmark-paper charts, method views, journal distributions, and gap maps from the shared paper set.

How The Workflows Connect

Skills share data through output/paper_set.json, so one workflow can hand off to the next without repeating the same searches:

/lit-review [topic]       -> builds the paper set
/lit-landscape [topic]    -> reads the paper set and generates figures
/brainstorm [topic]       -> tests idea novelty against the paper set
/idea [specific idea]     -> finds the closest papers from the paper set

Every search is also logged to output/search_log.md so you can see what the assistant actually looked up.

Optional Extras

Install as a Claude Code Plugin

The repository is packaged as a Claude Code plugin. Installing it adds the six skills, six slash commands, the paper-reader subagent, and the Corbis MCP server to any project in one step.

# From a Claude Code session, pointed at a local clone of this repo:
/plugin install ./

Alternatively, install directly from GitHub:

/plugin install github:Agentic-Assets/corbis-literature-starter-kit

The plugin manifest lives at .claude-plugin/plugin.json and references the skills and commands that already reside under .claude/. A parallel Codex manifest is provided at .codex-plugin/plugin.json for Codex users. Set the CORBIS_MCP_API_KEY environment variable before using any of the literature tools.

Project Structure

.claude-plugin/     Claude Code plugin manifest
.codex-plugin/      Codex plugin manifest
.claude/skills/     Workflow definitions and prompts (Claude Code)
.claude/commands/   Slash-command wrappers
.claude/agents/     Paper-reader prompt
.agents/skills/     Workflow definitions (Codex / other MCP agents)
notes/              Lab notebook
output/             Reviews, memos, figures, paper_set.json
latex_template/     Clean article template (natbib + plainnat)
utils/              Figure-generation helpers
references/         Writing norms and citation formatting

Research Insights

Journal-level publication stats, bibliometrics, and corpus overview in the Corbis research database (open in Corbis).

Open Datasets

When you need empirical data alongside literature work, Corbis hosts Open Datasets: a curated collection of (mostly) free finance research datasets you can search by topic, region, or use case. It lives in the Corbis product, not in this repo, but pairs naturally with /idea and related workflows.

Documentation