MCP (Model Context Protocol) is an open data exchange standard through which a language model or an agent built on artificial intelligence (AI) connects to external databases, file systems and applications. An MCP server is a software adapter that turns a structured request from the model into an executable command in an external information system.
The protocol replaces scattered custom integrations with one standard for connecting an agent to corporate data stores, terminals and application programming interfaces (APIs). When the standard was handed over to the Agentic AI Foundation under the Linux Foundation on 9 December 2025, there were more than 10,000 public MCP servers. The software development kits for Python and TypeScript were downloaded over 97 million times a month. The protocol delivers a call but does not assess how destructive its parameters are, so the check of an action before execution is placed between the agent and the server separately.
Why the protocol is needed
The protocol solves a connectivity problem. Before a single standard appeared, every connection of a language model to a corporate database, mail server or code management system required a separate connector written for that specific model. Whenever the model changed or the interface was updated, the whole integration layer had to be rewritten.
The protocol was developed by Anthropic, which on 9 December 2025 handed it over to a neutral foundation, the Agentic AI Foundation, under the Linux Foundation. It introduced a common client-server bus with three participants.
- The host application: the environment in which the user or an autonomous loop works, such as an agentic development environment, a corporate portal or an orchestrator.
- The MCP client: an internal module of the host that maintains a connection to a specific server over the standard protocol.
- The MCP server: an external or local adapter that announces a list of available functions and performs operations in the target system at the request of the client.
The move from closed connectors to an open standard brought rapid growth: by December 2025, the official software development kits (SDKs) for Python and TypeScript were downloaded more than 97 million times a month.
The Stacklok report “State of Model Context Protocol in Software 2026” was published in January 2026. It is based on a survey of 300 senior technology leaders at companies with 500 or more employees in the United States, Canada and the United Kingdom, in finance, retail and software development. According to the report, 29% of respondents were planning adoption, 30% were running pilots, 29% used the protocol in limited production and 12% in broad production.
What an MCP server is
An MCP server is a translator program that sits in front of a specific resource, such as a PostgreSQL database, a Git repository, a cloud console or a file folder. It offers the agent a strictly described list of operations. The server itself contains no artificial intelligence: it receives a structured call, performs the action in a trusted system and returns the result to the agent.
The protocol specification allows a server to offer the agent three kinds of primitives.
- Resources: readable data sources that are loaded into the context of the model, such as files, table rows, event logs and configuration schemas.
- Tools: executable functions with a description of their input parameters, such as
execute_sql,delete_file,create_containerandsend_payment; calling them changes the state of an external system. - Prompts: prepared scenarios and instructions for typical operations with the service.
To establish a connection, an administrator or developer gives the server the credentials of the target system: interface keys, database connection strings or service account keys. From that moment on, any agent that gains access to the server acts in the target infrastructure with the rights of that key.
How an agent calls a tool
From the intention of the model to a change in the database there are four rounds of message exchange in the JSON (JavaScript Object Notation) format under the JSON-RPC 2.0 protocol.
- Discovery: on connection, the client asks the server for the list of available tools, and the server returns the function names and a schema of the required arguments, such as the table name and the query text.
- Forming the call: analyzing the task and the list of tools, the language model produces not text for a human but a structured request to call a specific tool with computed arguments.
- Passing the command: the client forwards the request to the server over a local standard input and output channel or over the network.
- Execution and response: the server substitutes the arguments into a real system call or interface request, executes it under its own credentials and returns the result to the context of the model for the next step.
In this scheme the protocol guarantees that parameters are passed with correct syntax but does not check the meaning of the operation itself. If the model has formed a syntactically valid request to delete a production table or to export the customer register, and the key of the server has sufficient rights, an ordinary server will execute the command.
The specification says so directly: tools mean arbitrary code execution, and the protocol itself “cannot enforce these security principles at the protocol level”, so consent, access control and protection fall on those who implement it (specification revision of 28 July 2026).
| Stage | Direct interface call from code | Connection through an ordinary MCP server | Connection through MCP with an action control layer |
|---|---|---|---|
| Who forms the call parameters | A deterministic algorithm written by a programmer | A probabilistic language model, on the fly during reasoning | A probabilistic model, but parameters are intercepted before execution |
| Data structure check | At the level of programming language types and the interface schema | At the level of the protocol schema | At the level of the schema and a contract |
| Action safety check | Hard-coded into the script logic | Absent from the protocol itself; relies on the instruction to the model or the rights of the key | Check of the parameters by a solver under the contract before admission |
| Trace of the operation for review | Scattered entries in web server technical logs | An entry in the local log of the host, if enabled | A hash chain of receipts with external OpenTimestamps time stamps |
Why an MCP server without authentication is dangerous
An unprotected or over-privileged server becomes an open command gateway into the internal infrastructure of the company for any process or attacker able to send a network request. The server already holds working keys to databases and cloud environments, so without authentication at the entrance, privileged operations bypass perimeter protection.
The state of public servers was measured in early 2026.
Case: a survey of 1,412 public MCP servers (Bloomberry, February–March 2026)
Henley Wing Chiu of Bloomberry, in a publication of 24 February 2026 updated on 27 March 2026, analyzed 1,412 MCP servers publicly running on the internet. The data was later cited in a DreamFactory review of 3 March 2026. Among the servers whose authentication method could be determined with tools, 38.7% had no authentication at all, although the author did not disclose the exact size of this subsample. Only 2.4% of the servers surveyed limited the rate of requests, and 22.9% had a fully open cross-origin resource sharing (CORS) policy.
Even when a server is password-protected or runs locally on the machine of a developer, a second threat remains: prompt injection and the substitution of tool descriptions (protection against injections, page in Russian). If an agent reads an external document, a web page or a third-party repository with a hidden instruction, the model may itself call a local server connected to it. It may then execute a destructive command on behalf of a legitimate user.
The analysis of agent security incidents by Lumenova of 25 August 2026 (“Agentic AI Security Incidents”) records that standard pre-release tests detected none of the failures it examined. The root of the problems lay in how access to tools was arranged rather than in the quality of the models (“The failures were about access, not model quality”).
How to check agent calls before execution
A deterministic software control layer is placed between the client and the server, and it compares the parameters of each command with a contract of admissible actions. Checking a call with another neural network does not solve the problem, because the second model remains probabilistic and is exposed to the same distortions of context.
In the architecture of information security of the third generation (IS 3.0) developed at the institute, protection of the MCP loop follows the principle “check the action, not the intention” and is implemented in MITR.
- Intercepting the call: when the agent forms a request to a tool, the control layer intercepts the call before the command reaches the adapter and the target database.
- Checking against the contract: each class of actions has a contract in the form of a data file, and a class without a contract is refused by the solver. In enforce mode the refusal is enforced, while in advisory mode the agent receives a warning.
- Proof under the contract: the parameters of the call are passed to a satisfiability modulo theories (SMT) solver, which, without consulting a language model, checks whether the command can fall into the set of forbidden states. If it cannot, the control layer admits the call, and if a violation is found, a counterexample is produced together with the conditions for correcting the request. The proof holds relative to the contract, and the correctness of the contract itself remains with the human who approved it.
- Speed and accuracy: according to measurements of 24 October 2026 with 20 contracts in enforce mode, a decision under a rule takes 1.08 milliseconds at the median (1.96 milliseconds at the 95th percentile), and a call with an explanation takes 5.22 milliseconds. On 14 October 2026, with 19 contracts, the figures were 0.95 and 4.26 milliseconds. On an internal reference set of 157 cases (114 attacks on actions and 43 controls, prior to external accredited laboratory validation), MITR missed no attacks and produced no false refusals. The mode is set by a flag for each class of actions, and since 24 October 2026 all 20 contracts have run in enforce mode.
- Recording in a chain of receipts: each decision of MITR is written to a hash chain of receipts with external OpenTimestamps time stamps. In the historical epoch-33 anchor on 25 October 2026, the chain held 5,094 records and a replay by the checker written for an outside verifier matched in 5,079 cases out of 5,079. As of 15 November 2026, the live receipt chain holds 5,373 records across 55 completed epochs (56 published). For regulated Russian banks and critical information infrastructure, OpenTimestamps proofs are paired with accredited trust service time stamps and qualified electronic signatures under Federal Law No. 63-FZ. The institute itself carried out the replay, and no outside verifier has repeated it yet. The open kit at
asi5.ru/console/verify/lets anyone check the integrity of the chain and its time stamps independently.
What a manager should do
Safe use of the protocol on a corporate network requires five measures.
- Find deployed MCP servers. Check the external and internal network perimeters for public adapters launched by developers without authentication, with open cross-origin policies and without rate limits. In the public sample of Bloomberry, 38.7% of the servers with a determinable protection method had no password at all.
- Revoke full-rights keys from the servers. Move each adapter to service accounts with the minimum necessary rights, so that a server for reading analytics has no technical ability to drop tables or volumes.
- Ban unvetted third-party servers. Keep an internal register of approved MCP servers and review their source code before launching them in the development environment, to rule out the substitution of tool descriptions and the covert export of secrets.
- Put an action control layer in front of critical servers. Place a deterministic control layer in the call path, built like MITR, that checks the arguments of each command against contracts; with 20 contracts a decision takes 1.08 milliseconds at the median (measured on 24 October 2026).
- Keep a cryptographically linked log of decisions. Record all admitted and refused tool calls in a hash chain with external time stamps, which gives internal investigation and audit a verifiable evidence base; the architecture is described in detail in the book Proven Action. Power and Control over AI Agents: Strategy 2027–2035 by Aleksandr Fishchuk (2026, in Russian).
Frequently asked questions
Why is the protocol needed?
The protocol connects language models and agents to external data sources and working applications without a separate integration for each model. In December 2025 Anthropic handed it over to the Agentic AI Foundation under the Linux Foundation, when more than 10,000 public MCP servers were already running.
What is an MCP server?
It is an intermediary program that connects an agent to a specific database, file store or external service. The server announces the list of available commands to the agent, accepts structured requests from it and executes them in the target system under the credentials it was given.
How does an agent call a tool through MCP?
First, the agent asks the server for the list of functions and the format of their parameters. Then, while solving the task, the language model forms a structured command with the required arguments and sends it to the server, which executes the operation and returns the result to the context of the agent.
Why is an MCP server without authentication dangerous?
Such a server allows any external process or attacker to execute commands in the internal database or file system of the company with the rights of the key stored in its settings. According to the Bloomberry study of 24 February 2026 of 1,412 public servers, 38.7% of the servers with a determinable protection method had no authentication at all, and only 2.4% limited the rate of requests.
How can agent calls be checked before execution?
An independent software control layer is installed between the agent and the server; it intercepts each tool call and checks its arguments against a contract with a solver. In MITR, with 20 contracts in enforce mode, such a check takes 1.08 milliseconds at the median (measured on 24 October 2026) and does without a language model. It admits an operation only when its admissibility is proven under the contract.
Sources
- Anthropic: Donating the Model Context Protocol and establishing the Agentic AI Foundation. Official announcement of the company on handing the standard over to the Linux Foundation, 9 December 2025: https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation
- Stacklok: State of Model Context Protocol in Software 2026. Report on a survey of 300 senior technology leaders at large enterprises in the United States, Canada and the United Kingdom, conducted in December 2025 and published in January 2026: https://stacklok.com/wp-content/uploads/2026/01/State-of-MCP-in-Software-2026_FINAL.pdf
- Bloomberry (Henley Wing Chiu): an analysis of 1,400 public MCP servers. Study of 1,412 publicly running servers, 24 February 2026, updated on 27 March 2026: bloomberry.com/blog
- Lumenova: Agentic AI Security Incidents. Analysis of agent security incidents, 25 August 2026: https://www.lumenova.ai/blog/agentic-ai-security-incidents/
- Model Context Protocol: Specification, revision of 28 July 2026: https://modelcontextprotocol.io/specification/2026-07-28
- Measurements of MITR taken by the institute on 14, 24 and 25 October 2026. Open kit for checking the integrity of the receipt chain: https://asi5.ru/console/verify/
The protocol specification and the cited studies were checked on 16 October 2026; the MITR figures date from 24 and 25 October 2026.