Workflow and tool design
Define supported tasks, required inputs, data boundaries, and the actions that need user review. Start with one complete workflow that can be evaluated.
ChatGPT · Codex · Business integrations
ChatGPT Plugin Development & MCP Integration
Connect business tools to the tasks people complete in ChatGPT and Codex. TensorBlue helps scope plugins with reusable skills, controlled MCP tools, authentication, and optional app interfaces.
By TensorBlue Team · Updated
Choose the current plugin architecture
OpenAI’s current plugin model combines skills, MCP servers, and optional interfaces. Skills provide workflow instructions; MCP tools connect to live systems. A server can return structured results without a custom interface. Capabilities can vary by host and workspace, so we confirm the intended environment before implementation.
See the official plugin architecture documentation. If you arrive with an older ChatGPT plugin, we first inventory what it does and what the current platform supports. We do not carry forward legacy store assumptions or approval guarantees.
The useful starting point is a task such as “find this customer’s renewal date” or “prepare an update for these selected records.” Define the source system, account permissions, expected result, and failure path. This keeps the integration focused and gives the team a concrete acceptance test.
What the delivery scope can include
MCP server integration
Connect approved business APIs through narrowly scoped tools. Define input validation, useful results, errors, and server-side authorization.
Skills and packaging
Package reusable workflow instructions with the integration where useful. Specify when the workflow applies, what evidence it needs, and how to handle unsupported requests.
Optional app interface
Add a visual interface when people need to inspect records, compare options, or review a proposed action. Keep tool results useful when the interface is unavailable.
Evaluation and rollout
Check tool selection, arguments, account boundaries, failure handling, and end-to-end completion. Record release criteria and test the installed experience.
Operations and handover
Document deployment, credentials, monitoring, version changes, and support ownership. Plan how to recover from upstream API changes and revoked access.
From a business API to a dependable tool
Design each tool around a recognisable task. Use explicit identifiers and narrowly defined inputs. Return enough evidence to identify the source record and distinguish success, empty results, permission failures, and upstream errors. Keep credentials outside tool metadata and redact sensitive information from logs.
For an MCP server, OpenAI documents streamable HTTP and stable HTTPS endpoints for public submission. Private testing can use supported tunnel options. The production design must preserve authentication and authorization, and account for runtime support, streaming, upstream network access, logging, and rollback. See Build an MCP server for current requirements.
Our implementation review also checks retries, timeouts, and writes. A timeout must not silently create a duplicate order or record change. Decide how the user can confirm the affected record, how the backend detects repeated requests, and what evidence is returned after a write.
Evaluate the installed experience
Test direct requests, indirect wording, follow-ups, unsupported tasks, and actions requiring authorization. Record the chosen tool, arguments, returned evidence, errors, and review behaviour. Rerun the same cases after changing schemas or descriptions. OpenAI’s connection and testing guide describes capability testing followed by evaluation of the complete plugin.
For public distribution, prepare the package and supporting materials against the current submission requirements. Review decisions and timing remain outside TensorBlue’s control. A successful local test is a milestone, not proof that the published experience works for every account.
Example workflows to scope and test
Illustrative implementation ideas, not claimed client outcomes or measured performance results.
Sales operations
Find account records and prepare a proposed CRM update. Verify the selected account and changed fields before committing the write.
Evaluate correct record selection, authorized access, and duplicate-write protection.
Business analytics
Retrieve approved metrics and explain their date range and source. Preserve the distinction between retrieved figures and model-generated interpretation.
Evaluate metric definitions, source freshness, filters, and handling of unavailable data.
Customer support
Search approved policy material and draft a response for review. Route missing or conflicting evidence to a support specialist.
Evaluate evidence coverage, answer support, escalation, and reviewer effort.
Questions before commissioning a plugin
What does a ChatGPT plugin contain today?
Current OpenAI documentation describes plugins as packages of skills, MCP server capabilities, or both, with optional UI. We select the components from the workflow rather than require a server or interface for every project.
Can you migrate an older ChatGPT plugin?
We can assess the existing API, authentication, tool descriptions, and supported tasks. The migration plan should map those tasks to the current plugin model, then test the resulting integration. An old manifest or OpenAPI description alone does not establish compatibility.
Can it read private data or update records?
An MCP integration can expose controlled read and write tools. Access must be checked by the server for each operation, and consequential writes need appropriate review. Installing a plugin must not expand a user’s underlying account permissions.
Does public distribution guarantee approval?
No. We can prepare the package, testing evidence, endpoint, and submission materials. OpenAI controls review and publication; we do not promise approval or a fixed review time. Account and workspace policies also affect installation and use.
How much does development cost and how long does it take?
The estimate depends on API readiness, authentication, tool count, UI needs, deployment constraints, and evaluation scope. Discovery should separate the build from hosting, upstream service usage, and maintenance, and produce a phased proposal.
Do we need a plugin or a standalone AI application?
A plugin suits a workflow intended to run inside ChatGPT or Codex. A standalone application may be preferable when you need your own user experience, distribution, or operating controls. We review the target users and environment before choosing.
Bring a workflow, API documentation, and target environment
We can use those inputs to identify integration gaps, define evaluation criteria, and prepare a phased scope. Include the system owner and someone who can judge whether the task is completed correctly.
Request an integration assessmentFrequently Asked Questions
What architecture do current ChatGPT plugins use?
Current OpenAI documentation describes plugins that package skills, MCP servers, or both, with optional UI. We assess the target workflow and environment before selecting components or planning migration from a legacy integration.
How is data secured when ChatGPT calls our plugin?
The design should enforce authentication and per-user authorization on the server, restrict tool scope, and avoid exposing credentials or unnecessary personal data. Deployment and access policies depend on the target workspace and infrastructure.