TL;DR: Trigr runs one remote MCP server at https://trigr.xyz/mcp that lets ChatGPT, Claude, Claude Code, Codex and other MCP clients search market data, draft strategies, run point-in-time backtests and ML validation, and read full trade logs on your Trigr account. Every consequential action goes through a single-use authorization bound to its exact arguments, and the server has hard limits: no live trades, no execution starts, no withdrawals, no credential edits.
What is a trading MCP server?
The Model Context Protocol (MCP) is an open standard for connecting AI applications to external tools and data. A server publishes a list of tools with typed inputs; a client such as Claude or ChatGPT reads that list and lets the model call those tools during a conversation.
A trading MCP server applies that pattern to market research. Instead of asking a chatbot to imagine how a moving-average crossover "would have" performed, the assistant calls a real backtest engine and reads back real numbers. The difference matters: a language model with no tools will happily produce a plausible-looking equity curve that no engine ever computed.
In practice, a useful trading MCP server needs four things:
- Grounded discovery. The assistant must be able to look up which assets, timeframes, indicators and data sources actually exist, so it does not invent a feed.
- Real compute. Backtests and validation have to run on a proper engine, with costs and assumptions the assistant can report.
- Account context. The assistant should know your plan, credits and existing strategies before it proposes work.
- Guardrails. Tools that spend money, change saved work or touch trading need explicit, auditable approval, and some actions should be impossible from a chat window at all.
How does Trigr's MCP server work?
Trigr exposes a single remote endpoint, https://trigr.xyz/mcp, that speaks stateless Streamable HTTP. You add that URL to your client and sign in with your Trigr account in the browser. The browser flow creates a revocable MCP credential automatically, so there is no API key to copy into a config file. This follows the OAuth-based pattern described in the MCP authorization specification.
Each connected client gets its own credential and its own approval mode. Revoking one connection in Trigr's agent-access settings does not affect the others, and an account can hold up to five active MCP connections at a time.
MCP access is available on every Trigr plan, including Free. Normal account limits still apply: backtests and ML runs spend credits, and saved-strategy caps are shared across Studio, the REST API and MCP. Anything the assistant saves shows up immediately in Trigr Studio under the same account you already use, so you can open a draft the model created, inspect the node graph and keep editing by hand.
The server also ships three reusable prompts. In Claude Code they appear as slash commands: /mcp__trigr__build_strategy, /mcp__trigr__improve_strategy and /mcp__trigr__research_marketplace. In ChatGPT, Claude and Codex you can simply ask, for example, "Use Trigr to build a strategy." Each workflow starts by reading your plan, credits, strategy quota, saved strategies and agent summaries, then searches the capability catalog before proposing anything. Setup details for each client are in the guide to connecting Claude Code and Codex to Trigr.
What can your AI assistant do through Trigr?
The tools fall into three groups: read-only lookups, credit-metered compute, and account writes. The table below summarizes them.
| Group | What the assistant can do | Approval needed |
|---|---|---|
| Discovery | Search the capability catalog (assets, timeframes, indicators, data sources, parameter schemas); check point-in-time data coverage for an asset; search marketplace metadata and verified statistics | No |
| Account context | Read your plan, credits, quotas, saved strategies and trading-agent summaries | No |
| Stored results | Re-read stored backtest runs, including runs you made in the web app, and page through their full trade logs, for free | No |
| Agent activity | Read an owned agent's open positions and closed trade history, to compare what it did with its backtest | No |
| Compute | Run credit-metered point-in-time backtests with optional slippage and funding; start asynchronous ML optimization and poll the result | Yes |
| Strategy writes | Validate and save drafts; rename or replace an owned draft; publish, republish or delist an eligible strategy; delete an owned strategy | Yes |
| Agents | Create a paper agent in a paused state; edit an agent's name, capital or paused state | Yes |
A few details make these tools safer to hand to a model. Draft edits are guarded by a recipe hash, so if you change a strategy in Studio while the assistant is working, its stale edit fails instead of overwriting yours. Publishing, republishing and delisting also require the current recipe hash, and each one goes through the same exact-action authorization as a backtest; the usual publishing gates (paid plan, email, venue provenance and a stored verified full-history backtest) still apply. Backtests take a stable requestId so a retried call is not billed twice. Marketplace searches return statistics and metadata, but the node-graph recipe of a published strategy always stays private.
The assistant can also point a slot at an immutable marketplace version when it creates a paper agent, so the logic it tests cannot change underneath it. For the full list of tool arguments and the matching REST endpoints, see the agents and ML API reference.
What can't it do?
Some capabilities are simply not exposed over MCP, regardless of which approval mode you choose:
- It cannot start an agent's execution or place a live trade.
- It cannot withdraw funds or move money.
- It cannot read or edit API secrets, exchange credentials, signing keys or private risk configuration. Those values are never returned by any tool.
- It cannot change an agent's strategy or risk configuration.
The practical effect is a clear division of labor. The assistant does research, drafting and paper setup. A human, in the Trigr app, decides whether anything goes live. When you do deploy a live agent on Hyperliquid, it trades through a trade-only agent key that can place and cancel orders but cannot withdraw. The deeper reasoning behind these boundaries is covered in AI trading agent safety: approvals and hard limits.
How do approvals work?
When you authorize a connection, you choose one of two modes, and you can change it per connection later:
- Manual mode. Before each consequential action, such as a backtest, a draft save or a paper-agent creation, Trigr shows a confirmation in your browser. Nothing happens until you approve.
- Autonomous mode. The connection may continue without a prompt for each action. This suits long research sessions where you have already set a budget and a plan.
Both modes share the same mechanism underneath. The assistant first requests authorization for one exact tool call. Trigr issues a single-use authorization that expires after ten minutes and is bound to the exact arguments. If the model later tries to reuse it for a different graph, a different strategy or different cost settings, the call is rejected. Credit checks, quotas, ownership and graph validation stay active in both modes.
This design lines up with the MCP specification's guidance that there should always be a human in the loop able to deny tool invocations. Manual mode puts you in that loop for every action; autonomous mode keeps you in it at the level of the connection, which you can revoke at any time.
How do you keep AI-driven research honest?
An assistant that can run backtests quickly can also overfit quickly. Twenty variants in ten minutes will usually produce one that looks excellent by chance. Trigr's engine and its MCP tool descriptions push against that in several ways.
The engine is point-in-time. Every feature uses only information available at the bar's close, cross-asset inputs are carried forward rather than read ahead, and fills happen at the next bar's open after a signal. If take-profit and stop-loss are both touched inside one bar, the stop wins. The point-in-time backtesting guide explains why these choices matter.
Costs are explicit. Trading fees and the builder fee are always applied. Slippage and funding are opt-in and off by default, so a first result is gross. Over MCP, the assistant can add a flat slippage in basis points per side and funding, either the real historical series or a flat rate per 8 hours. The result reports what those costs were. Slippage is a flat assumption, not an order-book or market-impact model.
Overfitting is measured, not assumed away. An ML optimization run reports the Deflated Sharpe Ratio, which discounts a Sharpe ratio for the number of trials and non-normal returns (Bailey and López de Prado), alongside PBO and out-of-sample Sharpe. A standard backtest marks DSR and PBO as not yet computed rather than implying them.
The assistant is told what is not out-of-sample. A standard backtest uses the full available history and reports its last 30% as an out-of-sample split; the MCP instructions tell the assistant that this split stops counting as unseen once it informs a choice. For a final out-of-sample check, the assistant can start ML optimization, which selects among LightGBM, random forest and XGBoost by anchored, purged walk-forward out-of-sample performance, with no hyperparameter tuning, and returns a leak verdict, OOS Sharpe, DSR and PBO. Or it can create a paused paper agent so you can forward-test on data that did not exist when the strategy was chosen.
There is also a bounded backtest over a custom date window for detached research. It is useful for testing a frozen idea on a reserved period, and with the strategy's id the result is saved to Studio history labelled with its window. It never replaces the strategy's full-history statistics.
Backtests are not guarantees; perps are leveraged and can lose more than expected.
What does a typical research session look like?
A productive session with a connected assistant tends to follow the same arc, whichever client you use.
- Context. The assistant reads your plan, remaining credits and saved strategies, so it does not propose ten drafts on an account with room for two.
- Discovery. It searches the catalog for the exact building blocks: for example the 4H timeframe, an EMA indicator, and the funding-rate data source for ETH. It checks how far back point-in-time coverage goes.
- Hypothesis. You agree on one idea and write it down, including the costs you will assume and how many variants you are willing to test.
- Draft. The assistant validates the graph and saves one draft. You can open it in Studio and see every node.
- Test. It runs a gross backtest, then a net one with slippage and funding, and reports net return, drawdown and trade count side by side; DSR comes later, from an ML optimization run.
- Inspect. It pages through the trade log for free and flags anything suspicious, such as most of the profit coming from one week. Two units matter here: trade prices are bar reference prices before slippage (slippage is charged in the return, not the price), and each trade's P&L is in units of an account that starts at 100, not dollars.
- Decide. Either the idea is rejected, refined once with a written reason, or moved to a paused paper agent for forward-testing.
The assistant handles the tedious parts: exact ids, request formatting, reading hundreds of trades. You keep the judgment calls: which hypothesis is worth testing, what the trial budget is, and whether evidence is strong enough to act on. Wording for each step is collected in the AI strategy prompt recipes.
Which clients are supported, and what does it cost?
| Client | How to connect | Notes |
|---|---|---|
| Claude (web and desktop) | Add a custom connector with the Trigr URL | Custom connectors are available on Free, Pro, Max, Team and Enterprise; Free is limited to one |
| ChatGPT | Create a developer-mode app with the Trigr URL, following OpenAI's developer mode guide | Plan availability for write actions has changed over time; check OpenAI's current guide, and workspace admins may need to enable it |
| Claude Code | claude mcp add with the HTTP transport, then authenticate via /mcp |
Server prompts appear as slash commands |
| Codex | codex mcp add then codex mcp login |
Opens the browser authorization flow |
| Other MCP clients | Point at https://trigr.xyz/mcp |
Any client that supports remote servers with OAuth |
On cost, the model runs in your own assistant, so you pay your AI provider as usual. On the Trigr side, a standard backtest costs 50 credits, where 1,000 credits equal $1, and cache hits are free. ML optimization carries a surcharge of about 200 credits plus model cost. Reading stored results and trade logs is free. The Free plan includes 3,000 one-time trial credits and 10 saved strategies; Trader and Pro add monthly credits and higher limits.
Next steps
The fastest way to try it is the guided setup at connect your AI assistant, which walks through each client and lets you pick manual or autonomous mode. For a hands-on walkthrough of a first research session, read using ChatGPT to build and backtest a crypto strategy.