Skip to main content

running the algo optimizer from Claude, ChatGPT or the API

all access members can find templates, start and cancel optimizer runs, and read results from an AI assistant or their own code.

Written by Brad

if you're on all access, you can run the algo optimizer from Claude, ChatGPT or Grok through the edgeful connector, or from your own code with an API key. ask your AI to find a template, start a run, check on it, and walk you through the results, without opening the optimizer in the app.

this article covers: who gets it, the one setting to check, what you can ask for, how runs work from outside the app, and the limits.

who gets it

optimizer access through the API and MCP comes with an active, unpaused all access membership. essential and pro get the report API, but not the optimizer, and the optimizer tools come back with a permission error on those plans.

it uses the same connection you already have. if edgeful is connected to Claude, ChatGPT or Grok, there's nothing new to install. if it isn't yet, follow connecting edgeful to Claude, ChatGPT, and Grok with MCP first. for code, use your key from the API dashboard.

the one setting to check: writes

reading results works with writes off. starting or cancelling a run needs writes on. the switches are under let AI agents make changes in your account on the API dashboard: allow changes through the MCP connector for Claude, ChatGPT and other AI apps you've signed in to, and allow changes through your API key for your own code and coding agents. new or regenerated keys start with changes allowed.

if your AI can find templates and read results but says it isn't allowed to start a run, writes are off for that connection.

what you can ask for

you don't need the tool names. ask in plain English and the AI picks the right one. behind the scenes there are 6:

  1. describe_optimizer — the strategies, tickers and timeframes it accepts, a starting request that matches a fresh in-app optimizer, and the current limits

  2. list_optimizer_records — searches algo templates, or lists your own runs, saved results and favorites

  3. read_optimizer_result — the ranked results, then one result's settings, trades or charts

  4. submit_optimizer_job — starts a run

  5. get_optimizer_job — checks on a run

  6. cancel_optimizer_job — stops a queued or running run

it covers 7 strategies: ORB, ORB with 2 take profits, ORB with breakeven, IB, IB with breakeven, engulfing, and engulfing with 2 take profits.

some prompts that work well:

  • show me the ORB templates for MNQ and compare the top 3

  • run the IB optimizer on MES with holdout on, then show me the equity curve for the top result

  • list my saved optimizer results and tell me which one held up best out-of-sample

  • show the weekday breakdown and the trades for result #2

  • cancel my running optimizer job

start by searching what already exists. a template or an earlier run often answers the question without starting a new one.

how a run works from outside the app

  • it starts from the app's defaults. a new run begins with the same inputs as a fresh optimizer in the app for that strategy: one year of data, the full default sweep ranges and directions, and filters off. it doesn't use your saved in-app settings. tell the AI what to change.

  • holdout is off unless you ask for it. say "with holdout on" and it uses the default 80/20 split. the score works the same way as in the app, so read the out-of-sample numbers to judge how a setup held up (see understanding holdout, in-sample, and out-of-sample).

  • runs are queued, not instant. the AI starts the run and checks on it. you don't have to wait in the chat. come back later and ask for the results.

  • one run at a time, shared with the app. you get one active optimizer run, whether it was started in the app or from an AI. a run started from Claude shows up in the optimizer and can be cancelled there. if the app already has a run going, the new one waits.

  • an empty result tells you why. if nothing passes your filters, the response says which filters were out of reach and the best value any combination managed. loosen those first.

  • your AI draws the charts. equity, drawdown, monthly and weekday charts come back as data, and Claude or ChatGPT plots them. edgeful doesn't send an image. if your app can't draw charts, ask for a table.

limits

the search limits match the app: up to 20 million combinations (15 million for engulfing). on top of that, all your keys and connections share one budget for starting runs: one every 10 seconds, 60 an hour and 300 a day. a run that fails or is cancelled after it's accepted still counts.

runs started from outside the app have their own history of 50, separate from the app's. older ones are cleared after that unless you save or favorite them. if you hit a limit, the AI gets a "try again in N seconds" message and can wait it out.

from your own code

the same 6 actions are REST endpoints under /v1/optimizer, authenticated with your API key as a bearer token. the full reference, including how to retry safely and poll a job, is in the API docs: use the optimizer.

related articles

Did this answer your question?