Itoflow developer docs

1. Research. Ask in plain English from Claude, ChatGPT, Codex or your own code. An agent plans the work, writes and runs Python on Itoflow’s engine, and a second agent reviews it. You get the result with the code behind it: signal tests, walk-forward backtests, portfolio construction, and a factor risk model that looks through ETFs to their holdings.
2. Watch. Hand an idea, a list of stocks or a portfolio to a monitor: an agent that checks it on a schedule, under a mandate you write, and reports what changed.
3. Paper trade. Once you settle on a strategy, run as many paper versions of it as you want. Paper trades fill only while the market is open, at the latest available price, after a 0.10% commission and currency conversion. That price can come from an earlier session; see price timing. Each strategy records why it made every decision.
4. Go live. When a strategy is ready, move it to a live portfolio through Itoflow’s execution engine. A person approves every list of trades, and monitors keep watching the portfolio.

Every step happens in one place, so each starts from what came before: a live strategy carries its research, its paper record and its mandate. Itoflow also supplies what the agents need to run: prices, fundamentals, filings and news for stocks and ETFs on 24 exchanges in 22 markets, plus crypto (GET /symbols/exchanges lists them), and the compute. Start with one part. You don’t need the whole platform on day one. The risk engine works on its own over the API, so you can start by monitoring the portfolios you already hold, then add research and strategies.

What you can do

Published results

Each result states its conditions. The linked posts give the method and the limits.

• Factor risk model. Scored with the bias statistic and its 95% band as MSCI defines it, 0.82 to 1.18 for 59 month-ends: all 9 portfolio-family means and 107 of 116 portfolios fell inside the band, over 59 month-ends from September 2021 to July 2026 on a point-in-time S&P MidCap 400. On optimized portfolios the model came within 4% of realized volatility, where a plain sample covariance fell 17% to 26% short. One universe, one period. How our factor risk model measures up, September 8, 2026.
• Independent review. On 7 matched synthetic research tasks, one run each, review raised the mean score from 0.568 to 0.693 on a 0 to 1 scale. Six tasks improved and one got worse. How we use a second AI agent to review investment research, July 30, 2026.
• A run that found no edge. Asked to find a signal in US equities, the agent fixed a 12-stock large-cap universe before seeing any results, tested momentum and mean reversion out of sample, found no validated edge, and said so rather than invert the signals or switch universes. What our quant-research agent actually does — and where it still needs a human, July 9, 2026.
• Two sessions, end to end. Asked for a stock-bond strategy with a 10% volatility target and a 15% drawdown limit, the agent picked the only one of three designs that stayed inside the limit, which ran at about 7% volatility, and stress-tested it on the Covid crash. In the second, review sent a fossil-free MSCI World tracker back once before it passed. Both end in a paused paper strategy. Governed agents in practice, July 9, 2026.
• Tax-loss harvesting. Eight harvesting approaches and a no-harvest baseline on one modeled $10M S&P MidCap 400 account over five years: optimized sales with ETF replacement ended $1.08M (7.3%) ahead of no harvesting, with tracking error against the index of 1.56% a year, against 1.19% with no harvesting. Fixed rules and modeled fills; not tax advice. Finding tax alpha with Itoflow, September 29, 2026.
• News for each holding. On one test day of 25 stories, judged by an AI, rating news through the industries it could move found all 58 company-story pairs that rating every company found, at about a third of the cost. Matching company names found 4. It measured recall only: about one link in seven that it showed was one the judge marked. How we used Jev to find the news that moves a company without naming it, October 1, 2026.

BuySideBench is our public benchmark of buy-side tasks, with its method, conditions and results.

How the work is checked

• A checked plan before any work. Itoflow turns your request into a research contract: the question, the data, the dates, and the evidence the answer must carry. The system checks the plan before work starts. Each number the agent reports needs a proof the system checks: a computed value, a comparison, a saved file, or an out-of-sample choice that the system works out again itself.
• Overfitting checks in the engine. The engine computes the probability of backtest overfitting, the deflated Sharpe ratio and false-discovery-rate control, and the research contract can require them as evidence. The checks are code, not a prompt.
• A second agent reviews the work. Itoflow starts an independent reviewer on the plan before research and on the answer before it reaches you. The reviewer is a separate agent on the same model as the run. It reads the actual files, not the research agent’s summary, and either accepts the work or sends it back with what to fix. The research agent can’t skip the review or narrow it. If the review itself fails, or concerns are still open after 10 rounds, the work goes on, and the system’s own checks still apply.
• The engine starts the report; the agent’s analysis finishes it. A monitor report starts from the risk engine’s figures, and the agent may correct a figure or add its own from further analysis. A figure no one can produce shows as “not known”, never as a guess.
• Only data that existed at the time. A past-date risk read uses the factor model snapshot saved on or before that date, and Itoflow refuses a date or symbol the snapshot doesn’t cover rather than fill the gap. Fund holdings count from the day the SEC accepted the filing. Company statements count from when they became public, such as the results announcement, and the history reports any statement it couldn’t date that way.
• A person approves every live trade. Agents propose trades; they never send live orders on their own. Each approval covers one exact proposal, named by ID and a hash of its contents, so nobody can approve one list and send another.
• Every decision keeps its record. Chats keep their code, files and reasoning. Strategies store each decision with its weights, reasoning and confidence.

The agent harness explains how the checks work, and Design principles says what they mean for your code.

What Itoflow doesn’t do yet

• Stocks, ETFs and crypto only. Futures, options and individual bonds aren’t covered.
• One live broker. Live trading runs through Trading 212. Other brokers, such as Interactive Brokers, connect through their own MCP server, so agents can read your positions there.
• No live short sales. Long-short books, such as a 130/30, run on paper portfolios. Margin and financing are out of scope.
• No automatic live trading. Automatic execution works on paper portfolios only.
• Experimental APIs. All Itoflow APIs may change without notice. There is no official SDK: use any HTTP client, or generate one from the OpenAPI file.

Ways to use Itoflow

• From your AI assistant. Install the Itoflow plugin in Claude, ChatGPT, Codex, Gemini, GitHub Copilot or Cursor, sign in, and ask in plain English. Your assistant reaches the product through the Itoflow MCP server and never sees a token. Connect your assistant.
• In the Itoflow app. Run the same research, monitors and strategies in your browser at app.itoflow.ai.
• From your code. Sign in with device authorization and call the HTTP API with cURL, JavaScript, Python or any other client. Make your first request.
• With your own code, tools and models. Connect GitHub so agents can read, run and change your research code and open draft pull requests. Connect your own remote MCP servers so agents can call your tools. Tokens never reach the model or the sandbox. Chat and monitors can also run on your own ChatGPT subscription. Connect GitHub and MCP servers.

Your code, your data and your research stay yours. Itoflow sells no signals and imposes no investment view. Security. Connections use encrypted transport. Broker and ChatGPT-subscription credentials are encrypted at rest with AES-256-GCM. GitHub and MCP server tokens are held by our identity provider and never enter the sandbox. Itoflow is not a broker or custodian and holds no customer funds. Our SOC 2 assurance work is in progress. See the security overview. Plans. The free plan is enough to start. Pro and Max add advanced models and more included usage. AI usage counts in credits at the model provider’s cost times a stated multiplier. See Billing and credits and the pricing page. Itoflow is built by quants from J.P. Morgan and Tower Research and engineers from Google, Apple and Microsoft. The API coverage index lists every endpoint by feature. To ask an assistant about any page, choose Open in ChatGPT or Open in Claude from the page menu. To give it the whole docs, use the Markdown pages and llms.txt index.