gammarips-harness

gammarips-harness

An open-source agentic trading harness for the GammaRips options-flow data layer. Clone it, point it at the GammaRips MCP, and run a disciplined daily loop: your agent reasons over the overnight pool (liquidity-ranked, bullish only, one out-of-the-money call per name) to one trade candidate per day (or none), screens tradeability before thesis, designs its own exit, pre-registers every decision as data, and scores the whole pool after the close. Paper-only by default.

This is a workflow, not a signal service. GammaRips deliberately exposes no pick-returning endpoint, so there is nothing to copy. Two agents reasoning over the same pool, with different objectives and risk, should reach different contracts. The value is your analysis over shared data.

Try the data first (no key, no card, about ten seconds)

You do not need an account to see what this is built on. Point any MCP client at the anonymous tier:

claude mcp add --transport http gammarips https://mcp.gammarips.com/mcp

Then ask your agent for this morning's brief. The anonymous tier serves the pool preview, the daily report, regime context, the market calendar, and the methodology playbooks, with no signup. If the data is not interesting to you, you have spent ten seconds and you stop here.

The recording above is real output from the free tier on 2026-08-22. Nothing is staged. The run names no contract to buy, by design.

Install as a Claude Code plugin (two commands)

If you use Claude Code, install the harness as a plugin:

/plugin marketplace add DevDizzle/gammarips-harness
/plugin install gammarips@gammarips

The plugin bundles the GammaRips MCP server, the three loop skills (/gammarips:trade, /gammarips:review, /gammarips:coach), and the wiki-librarian agent. The bundled server is the anonymous endpoint, so the five free tools answer immediately with no key and no account. For the paid tools, export your key as GAMMARIPS_MCP_KEY and start Claude Code again, because the plugin sends that variable as a bearer token.

The skills write your dataset to eval/ and call scripts/, so run them inside a clone of this repo. The plugin gives you the server and the skills. The clone gives you the data directory and the scripts.

Running the harness (this part needs a key)

Be clear about which side of the line the harness sits on. Its screen is built on the paid tools, so the daily loop will not run on the anonymous tier:

Tool	Tier	What the loop uses it for
get_pool	free preview, pro full	the candidate pool; the screen needs the enriched view
get_daily_report	free	overnight market context
get_regime_context	free	the VIX term-structure rail
get_market_calendar_status	free	session timing
get_playbook	free	the methodology the agent reasons with
get_liquidity	pro	tradeability grading, which runs before any thesis
get_signal	pro	enriched per-name detail
query_outcomes	pro	scoring every pool name after the close
replay_contract	pro	exit-rule simulation

Tradeability grading and honest after-the-close scoring are the whole point of the loop, and both are pro. The harness is free; the data is what costs money.

1. Subscribe at gammarips.com/pricing ($29/mo founding price for the first 100, 30-day free trial).
2. Create your key at gammarips.com/account. It is shown once, so copy it then.
3. Export it:

   export GAMMARIPS_MCP_KEY="<your key>"
   

   .mcp.json reads this from the environment. Never hardcode the key.
4. Open the repo in Claude Code (or any MCP client that reads .mcp.json). The gammarips server connects over Streamable HTTP.
5. Run the loop:

   /trade    morning consult: screen the pool, rank 2-3 candidates or no-trade; you decide
   /review   after the close: score EVERY pool name, backfill the 3-day window
   /coach    on demand: behavioral feedback with receipts from your own record
   

   scripts/edge.py and scripts/skill_eval.py accumulate the evidence across days and say, with their own sample-size warnings, whether the screen is earning its keep.

What's in here

Path	What
CLAUDE.md	agent entry point: mission, non-negotiables, the loop
docs/TRADING-DOCTRINE.md	the rules: hard exclusions, tradeability, sizing (set your own), exit discipline
docs/OPERATIONS.md	MCP connection, timing, session procedure
.claude/skills/	the procedures: trade, review, coach, wiki-distill, ste100 (writing style)
.claude/agents/	wiki-librarian (wiki health)
.claude-plugin/	plugin + marketplace manifests: install this repo as the gammarips Claude Code plugin
wiki/	claim-tagged knowledge notes (findings/, literature/) + registries in _index/
eval/	YOUR dataset: funnel log, fills, behavior ledger, findings (starts empty)
scripts/	validators (funnel_log, behavior_log, lint) + evidence (edge, skill_eval)

Discipline

• Tradeability before thesis. The screen grades every name on early prints and book honesty before any story is considered. An untradeable winner is not a missed trade.
• Pre-register before outcome. One funnel row per pool name, written before any outcome is known; /review scores all of them, including everything the screen dropped. Scoring only your own picks cannot tell you whether the screen works.
• Cite claims with their maturity tag and exit-context. An edge is exit-conditional; a finding without its N is not admissible.
• The coach never advises. /coach referees your behavior against your own written rules, with receipts from your own record. It never says buy or sell.

Traces

traces/ holds one file per trading day of the harness running unattended on a VM, with Claude Code as the agent: the preflight verdict, the entry reasoning quoted verbatim, the data pulls and gate scripts it called, the monitor verdicts, and the ledger's entry, exit reason and fill. They are receipts, not claims. There is no win rate and no return figure in them by design.

Each file starts with two fixed lines: a notice that names what the file is (paper-traded research on a paper or shadow day, a live trading record on a live day, never investment advice), and the mode the day ran in (paper, shadow or live). The files are generated after the close by scripts/export_trace.py in the private operator repo, from the session transcripts and the trade ledger. Redaction is code, not a manual step: keys, tokens, account and order ids, balances, e-mail addresses, hosts and paths are stripped, a fixture test asserts none survive, and the export refuses to write a file in which any key prefix remains.

Not investment advice

GammaRips is a data vendor. Everything here is educational and on a paper-trading basis by default. Nothing in this repo is investment advice. Trade your own capital at your own risk, and get qualified securities and tax counsel before trading real money, especially if you publish signals you also trade.