Claude Code WRDS Toolkit (Liu fork of Orłowski's claude-wrds-public)

  • Author/Source: Siyang Liu, GitHub repo (fork of Piotr Orłowski's claude-wrds-public, with added IBES, Compustat, and Fama-French expert agents)
  • Original: https://github.com/lsy617004926/claude-wrds-public

  • Upstream: https://github.com/piotrek-orlowski/claude-wrds-public

  • Key Ideas

  • A drop-in set of Claude Code subagents and skills for autonomous WRDS work: copy agents/, skills/, and settings.json into ~/.claude/, append the bundled CLAUDE.md to your global CLAUDE.md, and Claude can query CRSP, Compustat, IBES, OptionMetrics, Fama-French, and TAQ from your terminal without the user writing SQL.
  • Seven specialist agents: crsp-wrds-expert, compustat-wrds-expert, ff-wrds-expert, ibes-wrds-expert, optionmetrics-wrds-expert, taq-wrds-expert, plus a wrds-query-orchestrator that coordinates multi-database queries (CRSP↔Compustat via crsp.ccmxpf_lnkhist, OptionMetrics↔CRSP via wrdsapps.opcrsphist, TAQ↔CRSP via wrdsapps.taqmclink).
  • Three skills: wrds-psql (connection patterns and single-line-command formatting rules), wrds-ssh (SSH connection, SAS batch submission, file transfer), and wrds-schema (a pre-loader skill --- /wrds-schema crsp optionm returns a compact reference card with table names, column types, date ranges, and known gotchas).
  • Permissions are pre-approved via settings.json for Bash(psql service=wrds*), Bash(ssh wrds*), Bash(scp * wrds:*), Bash(scp wrds:*). Without these, every WRDS call triggers a permission prompt.
  • Setup files required outside the repo: ~/.pg_service.conf (host/port/dbname/user for wrds), ~/.pgpass (with chmod 600), ~/.ssh/config (Host wrds entry pointing at wrds-cloud-sshkey.wharton.upenn.edu), and a one-time ~/scratch symlink (ssh wrds 'ln -sf /scratch/$(basename $(dirname $HOME))/$(whoami) ~/scratch') so all agents can refer to scratch via a portable path.
  • TAQ is special: too large for direct PostgreSQL, so taq-wrds-expert writes a SAS program locally, scps it to WRDS, submits via qsas, polls qstat, tails the log, and downloads the CSV. -sync y does not work on WRDS Cloud --- jobs are asynchronous only. TAQ scratch files auto-delete after 48 hours.
  • OptionMetrics tables are yearly (optionm.opprcd2024, not optionm.opprcd) --- a common pitfall the optionmetrics-wrds-expert is configured for.

  • Several design decisions worth borrowing for any WRDS-Claude setup: (1) use psql service=wrds instead of $WRDS_USERNAME shell variables, because shell substitutions trigger manual approval prompts every time; (2) enforce single-line psql commands (multi-line \ continuation also triggers prompts); (3) use psycopg2.connect("service=wrds") from Python rather than the wrds pip package, which uses interactive auth prompts that don't work in non-interactive agent contexts; (4) add SSH ControlMaster/ControlPersist 4h if Duo MFA is prompting on every TAQ-workflow step.

  • Summary

This is the canonical Claude-Code-friendly WRDS configuration --- the upstream that Dickerson's ai-asset-pricing repo (see summaries/dickerson-ai-asset-pricing) acknowledges as a starting point. The architecture is clean: per-database expert agents that know each schema's gotchas, lightweight skills that enforce shell-safe command patterns, and an orchestrator that handles multi-database link tables. The user just describes what they want ("Get me daily returns for AAPL and MSFT for 2024") and Claude routes the request through the appropriate agent → skill → SQL/SSH chain.

The most consequential design choices are about avoiding permission-prompt thrash. The pg_service.conf pattern, single-line psql enforcement, and pre-baked allow-listed Bash patterns all exist because shell-level dynamism (variables, multi-line continuations, command substitutions) interrupts agentic workflows with approval prompts that compound badly across long pipelines. This is a useful general pattern for anyone designing Claude tooling around any external service: the static, literal command form is the agentic-friendly form.

The Liu fork's contribution beyond Orłowski's original is the IBES, Compustat, and FF expert agents, plus the wrds-query-orchestrator. Liu's repo is the right starting point for someone setting up empirical asset-pricing or accounting work end-to-end.

  • Relevance to Economics Research

Direct and very high for this user. Mihail's primary data access is the WRDS Python package; this toolkit is the agentic-Claude analogue and would slot into any AssayingAnomalies-style setup. The IBES, Compustat, FF, and CRSP agents cover the cross-section-of-expected-returns workhorse stack. TAQ support via SSH+SAS matches Mihail's existing trading-cost / microstructure projects, and the asynchronous-only constraint plus the 48-hour scratch expiry are real operational details that would otherwise need to be discovered the hard way.

The borrowable patterns even for someone not adopting the toolkit wholesale: pre-approved-permissions in settings.json, the ~/scratch symlink for portable paths across users, pg_service.conf over shell-substitution, and the explicit "use psycopg2.connect("service=wrds") not the wrds pip package in agent contexts" are durable lessons for setting up any Claude-driven empirical finance project.