Connect ChatGPT or Claude to NoriaFlow with MCP

MCP connects an external assistant to NoriaFlow. After a one-time connection, a compatible assistant can inspect your workflows and use the tools permitted by your OAuth grant. The native AI co-pilot is in beta and appears separately in the editor only when the installation has a configured, reachable model; it does not require an external MCP connection.

What the AI can do depends on the scopes you approve. It can read drafts and node definitions, inspect the Discord structure your bot can see, and edit workflow steps and FLOW edges. It can also manage branches and saved versions, inspect saved values, use your image library, diagnose executions, request EvoGraph structural advice and run simulations. Publication and real runs have separate scopes and require explicit confirmation. The removed Memory/DATA wiring layer cannot be created, edited or enabled.

What MCP changes#

External assistants use MCP to read and act through NoriaFlow's tools. A typical workflow is:

1. Read the current workflow.
2. Look up the authoritative NoriaFlow node definitions.
3. Read bot-visible Discord structure when real server, channel, role, category, or emoji IDs are needed.
4. Validate the complete draft before saving it.
5. Apply the draft to an inactive workflow.
6. Run a safe simulation, inspect each node result, fix problems, and retest.

publish_workflow_branch is a separate operation that requires its own permission and explicit confirmation of the named branch. A dry test simulates external actions, so it does not send Discord messages or consume execution quota. Passing a test never implies permission to publish.

Prepare NoriaFlow#

1. Sign in to NoriaFlow.
2. Create or open the workflow the AI should build. It needs a saved workflow ID.
3. If the workflow is active, pause it before asking an AI to edit it.
4. For Discord workflows, connect your Discord bot and make sure it belongs to the server you want to use.
5. Keep the workflow editor open if you want to watch nodes appear and connect in real time.

Connect ChatGPT#

ChatGPT's current connector flow uses Developer mode:

1. Open ChatGPT → Settings → Security and login.
2. Turn on Developer mode. Availability can depend on your account or workspace policy.
3. Open ChatGPT Plugins, select the + button, and create a connection.
4. Name it Noriaflow and paste the endpoint shown above into the connection URL field.
5. Create the connection and review the tools ChatGPT discovers.
6. In a new conversation, open the tools menu and add the Noriaflow connection.
7. When NoriaFlow opens, sign in and review the requested permissions. Publication, Saved data, and image access are listed separately.

For the provider's latest interface details, see OpenAI's official connection guide.

Connect Claude#

For an individual Claude account:

1. Open Customize → Connectors.
2. Select +, then Add custom connector.
3. Name it Noriaflow, paste the endpoint shown above, and add it.
4. Select Connect, sign in to NoriaFlow, and review the requested permissions.
5. In each conversation, select the + button near the message box, open Connectors, and enable Noriaflow.

For a Team or Enterprise workspace, an owner first adds the web connector from Organization settings → Connectors → Add → Custom. Members can then connect it from Customize → Connectors.

For current plan availability and interface details, see Anthropic's official custom connector guide.

Build a workflow with AI#

When native AI is available, open AI beside View when the sidebar is collapsed. Describe the result you want, review the plan and proposed changes, then accept, reject or request a revision in the chat.

For an external assistant:

1. Open the target workflow in NoriaFlow. MCP in the topbar shows connection information.
2. Open your connected ChatGPT or Claude conversation and enable Noriaflow.
3. Describe the result and behavior you want, including rules and edge cases.
4. Ask the assistant to read the intended workflow and explain its proposed changes before applying them.
5. Approve the intended tool calls and review the resulting draft in the editor.
6. Wait for validation and the dry test to finish. Publish yourself, or explicitly approve publication of the named branch through the connector.

If the assistant cannot access the connector or a permission is missing, it should report that limitation instead of claiming that it changed your workflow.

You can also start from a normal ChatGPT or Claude conversation with a prompt like this:

Use the connected Noriaflow tools to build a Discord support-ticket workflow in my inactive workflow. Read the workflow and node definitions first, use my real Discord channels and roles, validate before saving, run a safe dry test, fix every error, and never activate it.

Permissions you approve#

NoriaFlow exposes separate permissions:

• workflow:read: read accessible workflows and NoriaFlow's node catalog, and request EvoGraph analysis or suggestions.
• workflow:draft:write: modify inactive drafts, branches and saved revisions; remove historical DATA links or confirm rollback to format 1. It cannot create/edit DATA links or authorize publication.
• workflow:test: launch safe simulations and inspect their redacted results.
• discord:structure:read: read servers, channels, categories, roles, and emojis visible to your bot.
• workflow:publish: publish a selected branch after explicit confirmation.
• memory:read: inspect your own runtime Saved data, with secrets redacted.
• memory:write: create, update or delete your Saved data after confirmation.
• media:read: search and inspect your own image library and usage.
• media:write: upload, rename and delete your images.
• workflow:run: start one of your active workflows for real after you confirm that run. Each run counts toward your monthly runs.

Review the permissions actually listed in the consent screen. A client that omits a scope list requests all currently supported scopes by default, including publication and real runs. A read-only grant does not permit edits or runs; each tool checks its own required scope. Your account's AI permission settings can restrict a grant further.

An existing connection keeps its original permissions. New permissions are never added to an existing token silently; reconnect when a new tool reports that its permission is missing. Workflow viewer/editor/owner rights still apply independently. Saved data is private to its workflow owner, even when the workflow is shared.

Bot credentials, secret node fields and private migration backups are excluded from tool results. Execution and Saved data reads apply the same redaction rules as the application. Real runs only happen through run_workflow, with the workflow:run permission and your confirmation. There are no credential-management, sharing-permission, billing or admin tools.

Run your workflows from ChatGPT or Claude#

Every active workflow with a Manual or Webhook trigger can become an action your assistant performs for you. Ask “What can you do in my NoriaFlow?” and the assistant calls list_runnable_workflows. Ask “Post tonight's schedule in my Discord” and it calls run_workflow with the input your workflow expects.

The first call only returns a preview of the workflow, trigger and input. The assistant must ask you, then call again with confirm: true. The run then happens for real: messages are sent, roles change, APIs are called. The assistant waits up to 30 seconds and reports the run status. When the workflow contains a Respond to Webhook step, its body comes back as the answer, so a workflow can look something up and return the result to the conversation.

List my runnable NoriaFlow workflows. Then run the one that announces a stream, with the title "Speedrun night". Show me the preview first and wait for my OK.

Ask EvoGraph for structural advice#

Ask the native chat or your connected assistant to “Analyze this workflow with EvoGraph” or “Suggest improvements and show me the changes before applying them.”

analyze_workflow_structure checks the graph, DATA links and available configuration context. suggest_workflow_improvements returns targeted recommendations and, when the snapshot is complete, conservative candidate diffs. The first version can suggest removing an isolated, disabled step that has no references or pinned output. Both tools are read-only and use the same adviser. Existing connections with workflow:read can use them without an additional permission.

The native chat currently sends partial metadata: some workflow settings and expression references cannot be checked through that projection. It receives diagnostics and recommendations with these limits reported; cleanup candidates and a complete total score remain unavailable. MCP can analyze the complete saved draft when the connected account has access to its full configuration.

The results explain what was checked and which information was unavailable. Scores describe static properties such as structural checks and simplicity; execution tests provide separate evidence about behavior. Review the reasons and proposed changes before applying them. These tools do not save, publish or execute a workflow, and they do not start learning experiments.

FLOW, historical DATA links and saved values#

FLOW edges schedule steps and choose control paths. Current workflows pass values to later steps through documented output references. New work uses schemaVersion: 1; a missing version also means format 1.

DATA links (dataLinks, format 2) belonged to the removed Memory wiring layer. They cannot be created, edited or enabled. For an older workflow, list_workflow_data_links can read existing links and delete_workflow_data_link can remove them from the draft. Neither action enables the old layer or publishes a change.

migrate_workflow_schema accepts only targetVersion: 1. It can roll back an older format-2 branch that still has its private migration backup. First read the workflow and preview the rollback with its latest baseUpdatedAt. Confirm only after reviewing that it restores the pre-migration graph and settings and discards later branch draft changes. Save a revision first if those edits need to be preserved.

The dashboard's Data page manages persistent saved values, a separate resource from these historical links. list_workflow_memory, create_workflow_memory, update_workflow_memory and delete_workflow_memory address saved values owned by the workflow owner. Updates/deletes use the entry's exact expectedVersion; system-managed, redacted and truncated entries cannot be edited. A graph revision restore does not restore saved values. See Working with data for the catalogue and storage workflow.

Branches and saved versions#

Technical schemaVersion, saved version numbers and branch names describe different things. currentBranch identifies the draft open in the editor; liveBranch and publishedVersion identify production.

Creating a branch does not switch to it. Autosave is separate from a manual commit. Historical revisions without a branch belong to main. Restore affects the draft and preserves its departing state by default; it does not republish. Publication runs NoriaFlow's normal validation and can enable commands, schedules and other real effects.

Workflow mutations use baseUpdatedAt from the latest workflow read. If another editor wins the race, the operation fails with a conflict. The AI must read and compare the current state before retrying, rather than blindly replacing the timestamp.

Images and diagnostics#

Use list_images with an optional filename search and cursor, then get_image to inspect an owned asset and its usage. The ID identifies the library record; the URL is the value to put in a workflow image field. Renaming keeps both ID and URL. Deletion is permanent and may break published workflows or historical revisions that still reference the image.

upload_image accepts actual base64 file bytes up to 512 KiB decoded. It validates and processes the image through the normal library service and deduplicates identical uploads for your account. It never downloads an arbitrary URL. Upload larger files through NoriaFlow's image library, then let the AI find the new asset using list_images.

get_workflow_diagnostics returns authoritative lint and recent node timings. list_workflow_activity, get_workflow_execution and get_workflow_execution_replay inspect real execution evidence without replaying effects. Use the dry-test tools and inspect their execution results to establish whether the workflow actually passes.

Confirmation and AI-readable documentation#

Publication, schema migration/rollback, version restore, runtime-data changes and image deletion first return a preview with confirmationRequired. The AI must show the operation and its consequences, obtain your approval for that exact action, and only then repeat it with confirm: true. The client handles this approval; the flag is not an independent verification of a human click. No special elicitation support is required.

The connector exposes the same operational guide at noriaflow://guides/workspace, through get_workspace_guide, and in the noriaflow-workspace-task prompt. These explain scopes, modes, versioning, image references and verification steps to the model. A preview is not a successful mutation, and an execution that failed is not a successful test.

Example requests:

Read the workspace guide and my workflow. Create an experiment branch, switch to it, and use the current node definitions to prepare a score workflow with storage steps and documented output references. Do not create DATA links or publish.
 
Find my welcome banner in the image library, use its URL in the image field, validate the draft, run a dry test, and save a manual version named "Welcome banner".
 
Inspect the failed runs from the last seven days, compare this draft with its previous saved version, and explain the cause before changing anything.

Troubleshooting#

The AI cannot find NoriaFlow tools#

Enable Noriaflow in the current conversation's tools or connectors menu. If it still does not appear, create a new conversation after reconnecting it.

A permission is missing#

Reconnect or re-authorize NoriaFlow and review the permission named in the error. Older authorizations do not include newer publication, memory or media scopes. Refresh the connection's tool metadata and start a new conversation after server updates; OpenAI documents this refresh flow.

The AI cannot see your Discord server#

Check that your bot is connected in NoriaFlow, is online, and belongs to that server. The connector only returns structure that the workflow owner's bot can actually see.

The workflow cannot be saved#

Pause active workflows before an AI edits them. If another editor changed the draft at the same time, ask the AI to read the workflow again before reapplying its work.

The connector is unavailable on your AI account#

Use NoriaFlow's native AI sidebar when it is available on your installation. It does not require an external MCP connection. The MCP dialog in the topbar manages external connections; workflow generation and preview review happen in the sidebar.

Setup checks and next steps#

• Check the draft before activating it; saving edits and publishing them are separate actions.
• Resolve missing parameters and connection errors in the workflow that reports them.

See the extension setup guide for adding and reusing services, the node reference for parameters, and testing and debugging for execution diagnostics.