Connect OpenAI to Claude: Build Assistants and Manage Fine-Tuning

If you need to connect the OpenAI platform to Claude to orchestrate fine-tuning jobs, audit usage costs, or manage project API keys, you need a Model Context Protocol (MCP) server. This infrastructure layer acts as the bridge, translating Claude's natural language tool calls into structured requests against OpenAI's REST APIs. If your team uses ChatGPT for internal operations, check out our guide on connecting OpenAI to ChatGPT or explore our broader architectural overview on connecting OpenAI to AI Agents.

Giving a Large Language Model (LLM) direct, programmatic access to your OpenAI organization is an engineering challenge. You are not just dealing with simple CRUD operations. You are orchestrating complex, asynchronous state machines for fine-tuning jobs, parsing highly specific usage data buckets, and navigating strict file upload schemas for batch workloads. Every time OpenAI releases a new endpoint or modifies a hyperparameter schema, your integration layer must adapt immediately.

Instead of building and hosting a custom integration layer, you can use Truto to dynamically generate a secure, authenticated MCP server. This guide breaks down exactly how to use Truto to expose OpenAI's administrative and operational endpoints to Claude, allowing your teams to manage AI infrastructure conversationally.

The Engineering Reality of the OpenAI API

A custom MCP server is a self-hosted translation layer. While the open MCP standard provides a predictable framework for models to discover tools, implementing it against the actual OpenAI API presents highly specific friction points.

If you decide to build a custom MCP server for OpenAI, you own the entire API lifecycle. Here are the specific architectural challenges you will face:

Complex State Machines and Polling OpenAI relies heavily on asynchronous processing. When you start a fine-tuning job, trigger a batch prediction, or execute an assistant run, you do not get the result back in the initial HTTP response. Instead, you receive an object ID with a status of queued or in_progress. An LLM cannot natively "wait" for an operation to finish. If you build this yourself, you must explicitly design your tool descriptions to teach the model how to poll these endpoints, check the status flags, and retrieve the final output files without getting trapped in infinite context loops.

Strict File Schemas for Async Workloads Uploading training data for fine-tuning or JSONL files for batch processing requires strict adherence to OpenAI's file validation rules. Files must be uploaded via multipart/form-data with exact purpose enums (e.g., fine-tune, batch, assistants). If a tool expects the LLM to format a multipart payload from scratch, it will frequently fail. A managed MCP server handles the underlying HTTP transport mechanics, allowing the model to simply pass the file content and purpose as standard JSON.

Usage Bucketing and Aggregation Extracting billing and usage data from the OpenAI API is complex. The /organization/usage/completions and related endpoints return time-bucketed data arrays that require start and end time parameters. For an LLM to effectively analyze this, the MCP server must cleanly expose these query parameters and map the resulting array of buckets into a flat, readable context window.

A Crucial Note on Rate Limits OpenAI enforces strict rate limits measured in Requests Per Minute (RPM) and Tokens Per Minute (TPM). It is important to note that Truto does not retry, throttle, or apply backoff on rate limit errors. When the upstream OpenAI API returns an HTTP 429 Too Many Requests error, Truto passes that error directly back to the calling LLM.

Truto does, however, normalize the upstream rate limit information into standard IETF headers (ratelimit-limit, ratelimit-remaining, ratelimit-reset). This explicitly shifts the responsibility of backoff and retry logic to the caller (your agent framework or Claude). Do not build workflows assuming your integration layer will automatically absorb rate limit spikes.

How to Generate an OpenAI MCP Server with Truto

Truto derives MCP tools dynamically from the connected integration's underlying resources and documentation. There is no manual tool coding required. As soon as you connect your OpenAI organization account to Truto, you can generate a secure MCP server URL.

You can provision this server via the Truto dashboard or programmatically via the API.

Method 1: Via the Truto UI

If you are setting this up for a local Claude Desktop instance or a quick internal workflow, the UI is the fastest path:

1. Navigate to the Integrated Accounts page in your Truto dashboard and select your connected OpenAI account.
2. Click the MCP Servers tab.
3. Click Create MCP Server.
4. Configure your access limits. You can restrict the server to specific HTTP methods (e.g., read-only access) or specific tool tags (e.g., only exposing billing and usage endpoints).
5. Copy the generated MCP Server URL.

Method 2: Via the Truto API

For enterprise environments automating workspace provisioning, you can generate the MCP server programmatically. Truto generates a cryptographic token backed by an edge data store, ensuring the resulting URL is fully self-contained and ready to execute immediately.

Send a POST request to the /integrated-account/:id/mcp endpoint:

curl -X POST "https://api.truto.one/integrated-account/{integrated_account_id}/mcp" \
  -H "Authorization: Bearer YOUR_TRUTO_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "OpenAI MLOps Server",
    "config": {
      "methods": ["read", "write"]
    },
    "expires_at": "2026-12-31T23:59:59Z"
  }'

The response will contain the secure URL you need to configure Claude:

{
  "id": "mcp-token-xyz",
  "name": "OpenAI MLOps Server",
  "config": { "methods": ["read", "write"] },
  "expires_at": "2026-12-31T23:59:59Z",
  "url": "https://api.truto.one/mcp/tkn_abc123..."
}

Connecting the MCP Server to Claude

Once you have your Truto MCP URL, you must connect it to Claude. Anthropic supports both UI-driven setup and configuration file management.

Method 1: Via the Claude UI

1. Open your Claude interface (Desktop or Web).
2. Navigate to Settings -> Integrations.
3. Click Add MCP Server.
4. Paste the Truto URL you generated in the previous step.
5. Click Add. Claude will instantly connect to the endpoint, perform the MCP handshake, and discover the available OpenAI tools.

Method 2: Via Manual Configuration File

If you are managing Claude configurations as code or orchestrating local agent environments, you can modify the claude_desktop_config.json file directly.

Because Truto operates as a remote HTTP-based server rather than a local standard-IO executable, you must use the official @modelcontextprotocol/server-sse transport package to bridge the connection.

{
  "mcpServers": {
    "openai_admin_mcp": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-sse",
        "https://api.truto.one/mcp/tkn_abc123..."
      ]
    }
  }
}

Save the file and restart Claude. The application will initialize the SSE connection and map the OpenAI endpoints into Claude's context window.

OpenAI Hero Tools for Claude

Truto dynamically translates OpenAI's endpoints into fully documented MCP tools. Claude uses the injected JSON schemas to understand exactly what parameters are required and how to structure the requests.

Here are some of the highest-leverage operations your AI agents can perform using this integration.

List Fine-Tuning Jobs

Retrieves a list of fine-tuning jobs across the organization. This is the entry point for monitoring model training pipelines, allowing Claude to track statuses across validating_files, running, succeeded, or failed states.

"Check our OpenAI account and list the 5 most recent fine-tuning jobs. Tell me which models they were based on and what their current status is."

Get Specific Fine-Tuning Job Details

Once Claude identifies a specific job ID, it can use this tool to retrieve granular details, including hyperparameters used, the trained token count, and the ID of the resulting fine_tuned_model.

"Get the details for fine-tuning job ftjob-12345abc. How many tokens were trained, and what is the final model ID?"

List Completions Usage

Auditing token costs is a massive pain point for engineering teams. This tool pulls time-bucketed token usage data, broken down by project, user, or model. Claude can use this to generate immediate spend reports.

"Pull the completions usage data for our organization starting from the 1st of this month. Summarize the total output tokens consumed across all projects."

Manage Vector Stores

Vector stores are required for OpenAI's Assistant API file search functionality. This tool allows Claude to dynamically provision new vector stores, which can then be populated with documents for RAG (Retrieval-Augmented Generation) workflows.

"Create a new vector store named 'Q3 Financial Reports'. Provide me with the new vector store ID once it is ready."

List Project API Keys

Security audits require constant monitoring of active credentials. This tool allows Claude to retrieve all API keys associated with a specific project, including their redacted values, creation dates, and the user who owns them.

"List all the API keys currently active in project proj-987xyz. Let me know if any keys were created more than 90 days ago."

List Organization Projects

Before managing specific users or keys, Claude needs to map the organizational structure. This tool lists all projects within the OpenAI organization, returning their names, statuses, and creation timestamps.

"List all active projects in our OpenAI organization. I need the project IDs to run an audit on service accounts."

For the complete inventory of available OpenAI operations, including batch processing, assistant thread management, and file handling, review the OpenAI integration page.

Workflows in Action

Exposing these tools to Claude enables complex, multi-step orchestration that would traditionally require custom Python scripts or internal dashboards.

Scenario 1: MLOps Fine-Tuning Monitoring

Machine learning engineers often trigger fine-tuning jobs via CI/CD pipelines but need an easy way to check status and clean up stuck runs without writing custom polling scripts.

"Check our ongoing fine-tuning jobs. If any job based on 'gpt-4o-mini' has been in a 'failed' status for more than 24 hours, retrieve its details so we can see what validation file was used."

1. Claude calls list_all_open_ai_fine_tuning_jobs to retrieve the array of recent jobs.
2. It filters the results in memory to find jobs where model matches gpt-4o-mini and status is failed.
3. Claude identifies a matching job and calls get_single_open_ai_fine_tuning_job_by_id using the retrieved job ID.
4. Claude reads the response payload, extracts the validation_file ID, and reports back to the engineer with the exact details needed for debugging.

Scenario 2: Platform IT Cost Auditing

IT administrators need to quickly identify cost anomalies and map token usage back to specific internal projects and users.

"I need an audit of our OpenAI token usage. First, list all our projects. Then, pull the completion usage for the last 7 days. Tell me which project consumed the most tokens, and list the active API keys for that specific project."

1. Claude calls list_all_open_ai_projects to get a mapping of project names to project IDs.
2. Claude calls list_all_open_ai_completions, passing in a start_time unix timestamp for 7 days ago.
3. It cross-references the project_id fields in the usage buckets with the project list to determine the highest consumer.
4. Claude calls list_all_open_ai_project_api_keys using the ID of the highest-consuming project.
5. The LLM presents a clean summary: The top consuming project, the exact token count, and the list of developers/keys responsible for the spend.

Security and Access Control

Connecting an LLM to your production OpenAI infrastructure requires strict governance. Truto provides multiple layers of access control at the MCP server level to ensure your agents operate securely:

• Method Filtering: You can configure the MCP server to only allow specific HTTP verbs. For example, setting methods: ["read"] ensures the agent can query usage statistics and list fine-tuning jobs, but physically cannot invoke create or delete tools, preventing accidental infrastructure modifications.
• Tag Filtering: Restrict tool access by functional boundaries. You can scope a server to only expose tools tagged for billing or fine-tuning, keeping the agent focused and minimizing the attack surface.
• Require API Token Authentication: By default, possessing the MCP server URL grants access to its tools. By setting require_api_token_auth: true, Truto forces the client to also provide a valid Truto API token in the authorization header, adding a strict second layer of identity verification.
• Time-Limited Access: Use the expires_at parameter to generate ephemeral MCP servers. This is ideal for CI/CD pipelines or temporary contractor access; once the expiration timestamp is reached, the server is automatically destroyed and the token is invalidated from the edge cache.

Moving from Dashboards to Agentic Infrastructure

Managing an enterprise AI platform via graphical dashboards is slow. Managing it via custom Python scripts requires constant maintenance against shifting API schemas. By connecting OpenAI directly to Claude via a managed MCP server, you turn natural language into an immediate execution interface for your infrastructure.

Truto removes the friction of pagination logic, complex asynchronous file uploads, and schema mapping. It exposes your OpenAI environment as standard, reliable tools, allowing you to focus on writing better prompts instead of maintaining integration code.