Read time … min

An MCP server for your trading runtime.

Bring your AI development tools into the TradingHost control plane. Inspect runtimes, follow logs, trigger redeploys and operate strategies from the assistant you already use.

Two pieces, that is the whole integration. A server URL at app.tradinghost.com/mcp, and an API key from your control panel. Drop the config block into your AI client, point it at the server, and TradingHost shows up as a tool the assistant can call.

Your AI assistant. Your TradingHost runtimes. One control surface.

MCP posture

  • Native MCP server
  • One JSON config block
  • Authorization Bearer header
  • API key from your control panel
  • Bring your own AI client
  • Control surface, not trade execution path
  • GitHub-driven deploy still applies

Section 01 · Position

MCP is the AI control surface for TradingHost.

Model Context Protocol is an open way for AI assistants to talk to external systems through a single declarative server entry. TradingHost runs one. Compatible AI clients connect to it the same way they connect to a file system, a database or a version control surface.

On TradingHost specifically, MCP is the route an AI assistant uses to operate your runtime. Triggering a redeploy after a commit, inspecting whether a strategy is alive, reading the last few lines of STDOUT after a connection error, restarting a process that has gone idle. The work the assistant requests happens inside TradingHost.

The control surface, the dashboard you log into, the GitHub-driven deploy flow and the MCP server all speak to the same control plane. MCP is the AI-native door into that plane.

SurfaceWhat MCP doesWhere it runs
DeploymentTrigger a redeploy, pick up a new commit, restart a strategy.TradingHost
Runtime inspectionRead deployment status, runtime metadata, region and lifecycle state.TradingHost
LogsStream STDOUT and STDERR from the running container.TradingHost
Trade executionOrder routing stays inside your strategy and the broker or exchange API your strategy already speaks to.Your strategy

MCP is how an AI assistant operates your TradingHost runtime. Order routing stays inside your strategy. The MCP server does not see your broker or exchange credentials, and it is not the path your trades take to a venue.

Section 02 · Connect it

One config block. One Bearer token.

Compatible AI clients accept a single JSON entry to register an MCP server. The TradingHost entry looks like this.

mcp.json
{  "mcpServers": {    "tradinghost": {      "url": "https://app.tradinghost.com/mcp",      "headers": {        "Authorization": "Bearer YOUR_API_KEY"      }    }  }}

  • Step 1

    Get your API key

    Open your TradingHost control panel and copy your MCP API key. The key is the value the snippet wants in the Authorization header. Treat it like any production secret.

    Open control panel

  • Step 2

    Drop the snippet in

    Paste the mcpServers entry into your AI client's MCP configuration. Most clients accept a JSON file. A few accept manual entry through a settings UI. Either path is fine.

  • Step 3

    Reload the assistant

    Restart or reload the AI client. It picks up the new server, opens a connection to TradingHost, and exposes the available tools to the model in your conversation.

No CLI. No installer. The integration is a config block and a Bearer token, the same shape every other MCP server expects.

If your AI assistant can already edit files, you can usually ask it to add the snippet to its own MCP config for you, paste the API key into the Authorization header, and reload itself. That is the fastest path on most clients today.

Section 03 · API key

Your MCP key lives in the control panel.

The Authorization header in the snippet wants a Bearer token issued from your TradingHost dashboard. The path is short.

  • Each key authenticates one AI client to one TradingHost account.
  • Keys are revealed once at creation. Store them in your password manager or secret store after that.
  • Rotate keys from the same screen if a laptop, browser profile or AI client is decommissioned.
  • Revoking a key disconnects every AI client that was using it. The strategy keeps running.

Section 04 · Compatible clients

Where the TradingHost MCP server connects.

The mcpServers entry is portable. Any AI client that supports MCP can register the TradingHost server with the same shape of configuration.

The clients below are the surfaces TradingHost expects most customers to start from.

ClientLaneHow it picks up the snippet
Claude DesktopDesktop assistantAdd the snippet to claude_desktop_config.json. Restart Claude Desktop and the TradingHost server appears in the tools menu.
Claude CodeCLI / IDE assistantAdd the server through the Claude Code CLI (claude mcp add) or drop the entry into the project-level MCP config so the agent can call TradingHost during sessions.
CursorIDEOpen the Cursor MCP settings panel and paste the mcpServers entry. The Cursor chat side-panel exposes the TradingHost tools to the model.
ChatGPT desktopDesktop assistantConfigure the TradingHost server through ChatGPT desktop's tools / connectors panel using the URL and Bearer token from the snippet.
CodexCLI / IDE assistantRegister TradingHost as an MCP server through the Codex configuration file. The Codex CLI and IDE companion both pick up the same entry.
VS Code with MCPIDEThe VS Code MCP integration accepts the same mcpServers JSON. Useful when an editor-bound agent needs to read TradingHost runtime output beside the code it edits.
Bring your own MCP clientAnything that speaks MCPAny client that can register an MCP server URL with an Authorization header should work. The protocol does the rest.

If your AI client is missing from this list, the integration is almost always still available. The MCP protocol is the contract, not the brand of the client.

Section 05 · What you can do today

The capability surface, by category.

The MCP server organises capabilities by what your assistant is trying to accomplish, not by HTTP endpoint. Today the server covers four areas.

  • Deployment

    Trigger and follow deployments

    Ask the assistant to redeploy a strategy after a commit, watch the deployment progress, and confirm the new version is running. The work happens through the same control plane the dashboard speaks to.

  • Runtime inspection

    Read runtime state

    Inspect deployment status, region, lifecycle state and metadata for a runtime. Useful when an agent is reasoning about whether a strategy needs attention before it makes a code change.

  • Logs

    Stream STDOUT and STDERR

    Pull the most recent log lines from a runtime. The assistant reads what the strategy is actually saying about itself, not a summary of the dashboard.

  • Lifecycle

    Operate strategy lifecycle

    Start, stop or restart a deployment. The assistant performs the same operations a human operator would perform from the dashboard, scoped to the runtimes the API key is allowed to reach.

The catalogue keeps growing as the server matures. Once an AI client is connected, the live tool list is visible inside the client itself, where it stays accurate without any marketing copy in the loop.

That is the source of truth for what is callable today.

Section 06 · Beta posture

Small surface on purpose.

The TradingHost MCP server is in Beta. The server itself is far from basic. The customer-facing surface is.

The integration is one JSON entry and one Bearer token. There is no installer binary, no plugin, no marketplace listing, no language SDK. That is intentional. The MCP protocol, the tool catalogue and the supporting AI clients are all moving fast right now, and a small surface lets us improve the server without breaking the way customers connect to it.

That is also why the marketing site does not enumerate exact tool names. The live tool list lives where it stays accurate: inside the AI client, after the server connects.

Still moving

What we are not pinning yet

  • The exact list of MCP tools the server exposes
  • How granularly an API key can be scoped to specific runtimes
  • How the server behaves under unusual load
  • How streaming logs are paginated for assistants
  • Which AI clients we actively QA against per release
  • How error responses are shaped for AI consumption

Stays stable

What you can rely on

  • The server URL, app.tradinghost.com/mcp
  • The Authorization Bearer header authentication scheme
  • The mcpServers JSON shape that registers the server
  • API keys are issued and revoked from the control panel
  • The control surface boundary, no broker or exchange credentials
  • The trade execution path, MCP is not on it

Honest framing

Beta means we keep moving the parts that are still maturing, and we hold the parts customers integrate against. If the server URL or the auth shape ever changes, it will change with notice. The mcpServers entry you copy today is the one you keep.

Section 07 · Boundary

The boundary, stated plainly.

The TradingHost MCP server is a control and management interface. It speaks for the control plane an operator already uses through the dashboard.

MCP does not see your broker or exchange API keys. Those values live in your secret store and reach the strategy as runtime configuration. The MCP server has no reason to read them and no surface to return them.

MCP is also not the path through which trades are executed. Order routing stays inside your strategy, against the broker, exchange or data provider your strategy already speaks to. An AI assistant requesting a "redeploy" or "restart" through MCP is asking TradingHost to operate the runtime. It is not asking TradingHost to place an order, cancel a position, or move a balance.

MCP authenticates through a Bearer token that customers issue and revoke from their own dashboard. Each key authenticates one AI client to one TradingHost account. If a key leaks, it is rotated from the same screen, and any AI client still using the previous value drops its connection at the next call.

That separation, control plane versus trade execution path, is what makes AI-assisted operations safe to lean into.

Control surface, not trade execution path.
MCP operates the runtime, your strategy operates the trade.

Section 08 · At a glance

The MCP surface in plain English.

QuestionAnswer
Is this an MCP server or an MCP client?A server. AI clients connect to it.
What is the server URL?https://app.tradinghost.com/mcp
How does an AI client authenticate?Authorization Bearer header with a TradingHost MCP API key.
Where does the API key come from?Your TradingHost dashboard, under settings, API keys.
Is this Beta?Yes. The customer-facing surface is small on purpose while the catalogue iterates.
Is the tool catalogue published?Not in marketing copy. The live catalogue is visible inside the AI client once connected.
Can MCP place trades for me?No. Order routing stays inside your strategy and your broker or exchange API.
Does MCP read my broker keys?No. Broker and exchange credentials live in your secret store, never on the MCP control surface.
Which AI clients are supported?Anything that speaks MCP, including Claude Desktop, Claude Code, Cursor, ChatGPT desktop, Codex and the VS Code MCP integration.
Do I need an SDK or a CLI?No. The integration is one mcpServers JSON entry plus an Authorization header.
What if my AI client cannot edit its own config?Edit the config file manually. Most modern clients also let an agent update their own MCP config when asked.
What happens if I revoke the API key?Connected AI clients drop on their next call. The strategy keeps running on TradingHost.

The important distinction is simple:

MCP drives your runtime. Your strategy still drives your trades.

Related

Where to go next

FAQ

Frequently asked questions

Yes. The customer-facing surface is intentionally small while the catalogue and the supporting AI clients keep moving. The server URL, the Bearer token authentication scheme and the boundary stay stable. The exact tool catalogue is the part still iterating.

No. MCP is a control surface for deployment, runtime inspection, logs and strategy lifecycle. Order routing stays inside your strategy and the broker or exchange API your strategy already uses.

No. Broker and exchange credentials live in your secret store and reach the strategy as runtime configuration. The MCP server has no surface that returns or accepts those values.

Drop a single mcpServers entry into the AI client’s MCP configuration. The entry needs a url of https://app.tradinghost.com/mcp and an Authorization header of "Bearer YOUR_API_KEY". Most clients pick up the new server after a reload.

From your TradingHost dashboard, under settings, then API keys. Reveal the value once at creation and store it the same way you would store any production credential.

Anything that speaks MCP. TradingHost expects most customers to start from Claude Desktop, Claude Code, Cursor, ChatGPT desktop, Codex or the VS Code MCP integration, but the protocol is the contract, not the brand of the client.

The catalogue is iterating week to week. The live list is visible inside the AI client once the server is connected, where it stays accurate without any marketing copy in the loop. That is the source of truth for what is callable today.

No. The integration is one JSON entry plus a Bearer token. There is deliberately no installer binary, plugin or marketplace package while the server is in Beta.

Edit the configuration file manually. On most modern clients, the MCP config is a small JSON file you can open in any editor. Many clients also let an agent update their own MCP configuration on request, which is often the fastest path.

Connected AI clients drop on their next call. The deployed strategy keeps running on TradingHost. Issuing a new key from the control panel and updating the snippet restores the connection.

Connect your AI assistant. Drive your runtime.

Grab your MCP API key from the control panel, drop the snippet into the AI client you already use, and bring TradingHost into your workflow as a tool the assistant can call.

MCP drives your runtime.
Your strategy still drives your trades.