Bring your own agent.

Connect Claude, ChatGPT or any MCP agent and it can read what you see in NYTSHIFT and help when something is confusing: why funding is high, what a screen tier means, how big a coin really is. It starts with market data only. You decide whether it may also read your accounts or paper trade, and you can test, rotate, pause or remove it at any time in Agents.

Limits

• 120 calls a minute for each key, and 240 a minute for all of your keys together.
• 30 candle or order-book reads a minute for all of your keys, because those can reach the exchange. Use the market list or the screen for many markets at once.
• An agent over a limit is told how many seconds to wait, and answers to a working key carry RateLimit headers so an agent can pace itself.
• Up to five active keys or sign-ins. A key lasts 7, 30 or 90 days.

Test, rotate and remove

• Test connection sits next to a new key, and under Connected agents you can test a saved key. It checks the key, the tools, the permissions and one live price.
• Rotate key gives the same agent a new key with the same permissions and lifetime. The old key stops at once. ChatGPT and Claude sign-ins renew themselves every hour and are removed instead.
• Remove ends a key or sign-in on its next call. Pause all agents stops every tool call until you resume.

If something does not work

• Key not accepted (401). The key is wrong, was removed or rotated, or has expired. Create or rotate a key in Agents and paste the new one.
• Paused. All agents are paused on the Agents tab. Resume them there.
• Too many calls (429). Wait the number of seconds the answer gives. Ask broad questions with the market list or the screen instead of reading markets one by one.
• Needs Agent Pro. Market data stays free. If the site has switched to Agent Pro, reading accounts, paper trading and live requests need it.
• No Developer mode in ChatGPT. Your ChatGPT plan or workspace does not offer custom connectors. Use Claude or a key-based agent instead.

What a connected agent can use:

• Markets. Available prices, volume, native funding, open interest, Radar, candles and order books retain their venue and original source clocks. Unsupported or stale evidence stays unknown.
• My accounts (only if you share them). Share exact Hyperliquid accounts or verified native account UUIDs for Lighter Core, Lighter on Robinhood Chain and Arcus. accounts:read covers those selected accounts, not every wallet or account. Native balances, positions, orders and precision come from checked native reads; missing USD values, fees or protection remain unknown.
• Playbooks. Grant private research and library access separately. An agent can read the guide, validate rules, backtest an exact revision and save it. Saving or backtesting never starts a run. Server PAPER runs require the separate run-paper grant, paper:trade and current configured availability.
• PAPER trading. Explicitly initialize simulated capital, then place simulated orders through the existing PAPER engine. Agent PAPER fills and retained history appear in Trade and Journal. Reset starts a new account generation and preserves archived evidence.
• Journal. accounts:read includes saved Journal records and private notes within the shared account scope. Native Journal tools preserve physical account identity, actual fill IDs, units and bounded coverage; incomplete costs never become a complete net result or an invented R multiple. Unsaved browser observations stay local.
• Guardian. Separate owner-granted read/write access lets the agent review or manage alert rules through alert_list, alert_create and alert_delete. Guardian cannot trade. Account alerts use current observations while the app is open; configured market alerts can run on the server.
• Market Lab theses. Separate private thesis read/write grants allow thesis_create, thesis_list and thesis_get. Saved plans retain their author and revision; a thesis does not authorize an order.

Granted tools can save private research, notes, theses and alert rules, or change simulated PAPER money. They do not create their own access or change your sign-in. Shared account records and saved notes are sent to the AI provider behind your client when it reads them; choose the account UUIDs and grants accordingly.

Live actions with your approval. Where configured and supported by the selected venue, orders:request allows a typed order, position or order-control request within the limits you approved. The pending request shows the native operation, account, terms and reason in NYTSHIFT. Nothing is sent until you approve that individual request; unanswered requests expire and uncertain outcomes are reconciled rather than automatically retried. An explicitly owner-armed Playbook can file requests after its exact backtest and PAPER evidence are checked. It never approves or executes them.

• MCP is not a signer proxy. A trading request is a pending proposal, not a direct venue submit or cancel. Research and PAPER tools cannot sign, send, transfer or withdraw real funds.
• Public addresses still require an explicit account grant, because sending account context to an external model is a privacy boundary.
• Loopback HTTP and stdio belong to the separate legacy local bridge and its operator setup. Hosted customer MCP and supported OAuth sign-in use customer permissions, current account ownership and server availability; local operator authority is not a prerequisite or a customer grant.
• NYTSHIFT never asks for OpenAI, Anthropic, Google or other model-provider API keys.
• Plan or payment state cannot grant execution authority.

Long-form detail lives in the Bring Your Own Agent chapter ↗ of the guide.