T-Investments has launched a public MCP server that connects AI agents and automation platforms to its brokerage API. Compatible clients can search for securities, inspect portfolios and market data, and—when granted the required permissions—place or cancel orders and initiate supported transfers.
The remote server uses Streamable HTTP, with separate endpoints for live accounts and a sandbox containing virtual accounts. Authentication relies on the same bearer token used by the T-Investments API. When issuing a token, users select an account and choose read-only, trading or transfer permissions; a sandbox-only token is also available.
According to T-Investments, the MCP tools mirror its public API and add service functions. They cover positions, available cash, portfolio valuation and returns, transaction history, order books, candles, recent trades and technical indicators. Other functions include instrument search, limit and market orders, stop-losses, take-profits, news, fundamental data, dividends, coupons and bond events.
The company presents two implementation examples. Hermes Agent can power a scheduled Telegram assistant that checks whether portfolio holdings lack take-profit orders. An n8n workflow provides a more deterministic approach: it can rebalance an account each evening between a money-market fund and Russian federal government bonds, while requesting trade confirmation through Telegram.
Reproducing the examples requires Docker, a Google AI Studio API key, a Telegram bot token and a T-Investments token. The environment must also contain certificates issued by Russia’s Ministry of Digital Development. T-Investments recommends starting in the sandbox, which offers the same tools as the live environment but cannot access real accounts.
T-Investments advises creating a separate token for each scenario, granting only the minimum necessary permissions, and keeping secrets in environment variables or encrypted credential storage. Access to the agent itself should also be restricted. If an agent receives trading privileges, the developer must deliberately decide whether a person should remain in the approval chain; an informational assistant does not need trading access.
Practical context: MCP can reduce the amount of custom integration code, but it does not make autonomous trading deterministic. An agent may be useful for conversational or varied requests, while its chosen sequence of tool calls can differ between runs. For recurring financial operations, the source supports a hybrid design: a fixed workflow enforces checks and constraints, while an agent handles only the steps that require flexibility.
| Criterion | AI agent | Automation platform |
|---|---|---|
| Logic | The LLM chooses the sequence of calls | The author defines a fixed graph of nodes |
| Interface | Natural language | Visual editor |
| Predictability | The plan may differ for identical inputs | The same data follows the same route |
| Changing behavior | Rewrite the prompt | Rebuild the workflow |
| Suitable tasks | Conversation, research and varied requests | Recurring processes where reliability is critical |
Sources
Event date: 2026-09-26. Primary source date: 2026-09-26.
