cairn
A local-first MCP server that connects Claude Code to your Robinhood account, enabling read-only portfolio insights, tax analysis, and natural language queries, with optional human-in-the-loop trading actions.
README
Cairn
An insights engine for your Robinhood account that runs on your machine. It unifies everything Robinhood scatters — live positions, trade & options history, and monthly statements — into one report you can actually read and ask questions of: concentration, cost basis, holding period, realized long/short-term gains, premium income, wash-sale flags, and opportunities.
Robinhood keeps your data on three surfaces that never talk to each other — the UI (now), history (the past), statements (the record) — and none of them synthesize. So the insights you actually need are impossible to get without exporting CSVs and hand-stitching spreadsheets. Cairn does the stitching, on your machine.
Built in the open. Runs local-first — your credentials live only on your device, the connection is strictly read-only, and every line is auditable. MIT licensed.
⭐ If Cairn shows you something Robinhood couldn't, give it a star — it's the only way I know it's useful to someone.
Quick start
git clone https://github.com/arunash/cairn.git && cd cairn
cp .env.example .env # add your Robinhood login — stays on this machine
pip install -r requirements.txt
python -m cairn report # builds web/report.html from your live account
open web/report.html # your insights, on one page
Then open the folder in Claude Code and just ask: "what should I do about my concentration?", "what's my tax situation?", "how much premium did I collect?"
You need nothing but a Robinhood login for the report. For the chat you need Claude Code (it's the brain). Everything runs locally.
Where does it open?
web/report.htmlis a plain file on your machine —openit (file://), or runpython -m cairn serveto build it and open it athttp://localhost:8787(bound to localhost only; add--port Nto change). There is no hosted portal — your data never leaves your machine.
Status — built in the open. The report spans all your accounts and is a full read: value, cash, every position with cost basis / unrealized P&L / holding period (LT vs ST), concentration (Herfindahl index), realized long- and short-term gains, options premium, dividends, and wash-sale candidates. Cost basis comes straight from Robinhood (already split- and transfer-adjusted); realized gains use a small bundled, local split table — no network.
python -m cairn taxesprints the tax view, and--account <id>limits any command to one account. Read-only throughout. Issues and PRs welcome.
The three layers — all free, all local
| Layer | What it does | Runs on |
|---|---|---|
| 1 · Insights | The report — deterministic compute + render | Python, no LLM |
| 2 · Ask Cairn | Chat & advice, in the context of your whole account | Claude Code, read-only |
| 3 · Act & Steward | Places trades, runs the wheel, rolls, trims, watches | Claude Code + write tools, human-in-loop |
The safety line is structural: layers 1–2 only ever have read tools, so they physically cannot trade. Layer 3 unlocks the write tools and still approves every single order with you.
How it works
CLAUDE CODE (the brain, local) YOUR ROBINHOOD (your account)
┌───────────────────────────┐ read ┌────────────────────────────┐
│ CLAUDE.md = rules+guards │ ───────────► │ positions · quotes │
│ config/ledger.json = YOU │ ◄─────────── │ trade & options history │
│ broker/mcp.py = the gate │ write* │ monthly statements (PDF) │
└───────────────────────────┘ *Act only └────────────────────────────┘
│ runs
▼
cairn/ report ──► web/report.html python -m cairn report
compute/* · act/* · watch/* gains · premium · wheel · roll · trim · alerts
- Report (
python -m cairn report) pulls your live holdings read-only and renders the page — account value, unrealized P&L, concentration, and per-position cost basis & weight. - Compute holds the deterministic engines in
cairn/compute/: a FIFO lot engine (realized LT/ST gains, holding periods, wash-sale flags), positions & concentration, premium, cashflow (deposits + dividends), and the intent ledger (pool split, day-over-day). - Claude Code reads the report, answers questions in plain language, and — only in Act mode — executes with your approval.
Security
- Local-first. Your Robinhood login sits in
.envon your machine. There is no Cairn server. - Read-only by default. Insights & chat never have trading permission.
- No outbound calls except to your Robinhood account and (for chat) your Claude.
- Disconnect anytime — delete the local session token and it's cut off instantly.
- Auditable — it's all here, MIT, nothing obfuscated.
Repo structure
cairn/
├── CLAUDE.md # the agent's behavior, YOUR rules, the hard guardrails
├── SECURITY.md # read-only-by-default model + secret handling
├── .mcp.json # Robinhood MCP wiring (read tools always; write gated to Act)
├── config/
│ └── ledger.example.json # intent-based pools + your rules (copy → ledger.json)
├── cairn/
│ ├── broker/mcp.py # the MCP gate — write tools registered only when CAIRN_ACT_ENABLED=1
│ ├── report/build.py # builds web/report.html (the insights page)
│ ├── compute/ # lots · realized · washsale · positions · premium · cashflow · ledger
│ ├── ingest/statements.py # monthly statement PDF → ACH cash-flow parse
│ ├── act/ # wheel · roll · reinvest · trim (human-in-loop)
│ └── watch/ # trim watcher · reminders (email alerts)
└── web/report.html # generated locally — your insights page (git-ignored)
Keeping your fork clean
Your personal data never belongs in the repo. Account IDs, cost basis, credentials, and generated output all live in git-ignored files only:
| File | What | Where it comes from |
|---|---|---|
.env |
your Robinhood login | copy .env.example |
config/ledger.json |
your accounts & pools | copy config/ledger.example.json |
web/report.html, config/daily-ledger.* |
generated output | created at runtime |
The tracked cairn/ modules ship with placeholders only — MARGIN_ACCT, IRA_ACCT,
<ticker>, BASIS = 0.0. Put your real values into config/ledger.json, never into the code.
Before you push, confirm git status shows none of the files above. See SECURITY.md.
License
MIT — do anything, no warranty. This is software, not financial advice. You run it, you own the decisions.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.