Find and Execute a Connected Operation
For a known organization, a simple external-app lookup takes two MCP calls:- Call
workspace.searchwithorganizationId, an operationquerysuch asnote, andconnectionQuerysuch asAttio. - Fill the caller inputs described by the matching result’s
executionfield and call itsinvocation.toolNamewith its fixed arguments and those inputs.
connectionQuery matches authorized connection display names and provider
identifiers using case-insensitive literal matching, independently of the
operation query. It intersects with connectionId and capabilityKind.
Duplicate connection names remain separate results with connection names and
Space exposure metadata. Select the intended connection explicitly.
Capability search paginates matching operations across the authorized connections.
A successful empty result means the search is exhausted. If a search exceeds its
scan budget or deadline, it returns an error. To enumerate a connection’s operations,
read its capabilitiesUri and follow the catalog’s resource enumeration links.
Executable definitions include input schemas and their placement, command tags
and version pins, or named authenticated-fetch operations with recipe-owned
business input schemas. For an operation read, use kind: "operation", its
operationId, and put values from inputSchema under inputs. For a native MCP
read, use kind: "mcp" and put values from inputSchema under arguments.
Neither shape accepts caller-supplied HTTP methods, paths, or request templates.
Unavailable operations explain why direct execution is unavailable. Each response
includes at most 64 KiB of execution definitions; pagination preserves complete
schemas. A single oversized definition returns its exact resource reference
instead. Existing capability resources retain complete YAML, Python examples and
other workflow-authoring material.
Bundled and uploaded recipe commands use the same capability resources. A command
child resource includes inputSchema, outputSchema, safety metadata, and
availability. commandOutputPath identifies the result described by outputSchema.
Copy its commandInput, fill commandInputPath with your inputs, and add the current
organizationId when calling connections.command. Uploaded commands include their
revision and operation pins, as described by the
uploaded command data schema.
Flowglad resolves credential ownership and the connection’s Space internally. Commands marked
requires-workflow need an authorized workflow with the required file context.
connections.command executes immediately and the call counts as approval.
The originating Automation’s auto-execution setting governs automatic approval of workflow command proposals. Current permissions, feature
gates, operation admission and version pins are checked again at execution.
Provider-supplied structured fields and resource links pass through command
results; no automatic follow-up enrichment reads are performed. Operations
that require workspace-specific attribute discovery still need that discovery.
Set Up a Connection
Organizations enabled for connection setup expose five additional tools:
Creation requires the exact catalog product, an explicit Space, an
account choice
(authorize or a reusable account), and a stable requestKey. Bundled product
slugs distinguish products that share a provider; uploaded references pin their
definition and revision. On the universal endpoint, include organizationId
when selecting among authorized organizations.
An authorization_required result contains an app-origin continuationUrl.
Open it in a browser signed into the same Flowglad account that owns the MCP
grant. Another organization member or administrator cannot complete the handoff.
Enter API keys, passwords, client credentials, service-account material or static
MCP credentials only in that browser form; OAuth consent also runs there. Never
put credentials, headers, provider tokens or credential JSON in tool arguments.
The link identifies the attempt and does not convey bearer authority.
After authorization, account and resource selection stays with the agent. Scope
discovery supports the product’s declared resources, filters, folders, workbooks
and worksheets, provider accounts, or remote MCP operation review. Configure
using the current expectedRevision; remote review also requires its discovery
revision. Stale or invalid selections conflict without replacing saved scope.
get never writes. discover preserves selected scope, although explicit
refresh: true can persist provider discovery and requires management consent.
connectors.list, connections.accounts.list, and scope discovery paginate
eligible results. They consume unavailable candidates and empty provider pages
internally. A successful empty result means the eligible inventory is exhausted;
discovery_incomplete reports that a scan or time budget prevented completion.
Narrow the query or resource scope before retrying an incomplete discovery.
Setup activates automatically once authorization, saved selections and provisioning
are ready. status only reports progress; no public completion action is needed.
Pending configuration cannot execute as an active connection. Equivalent usable
product/account/Space/scope requests return the same connection. A scope mismatch
requires an explicit configure operation and never silently broadens access.
An unchanged requestKey resumes the same attempt; changed immutable intent
conflicts. retry resumes unfinished work after a recoverable failure. Closing
the browser leaves an expiring, resumable attempt. cancel discards only unfinished
setup and preserves reusable authorizations and active connections. Expired
attempts require a new creation request.
Reads require delegated connection read permission and current Space access.
Creation, retry, cancellation, browser authorization, configuration, persisted
discovery refresh and activation require management authority. The consent screen
offers connections:manage and space:connections:manage; existing read-only
grants remain read-only until explicit reconsent. Current grant, owner, organization,
Space, product restrictions and feature availability are rechecked during setup.
Both mixed-action tools advertise write/destructive capability, even though their
read actions are observational. Failures return bounded safe codes rather than
provider diagnostics or credential material.
Call Tools
Agents can discover, enable, and call published provider operations without a Connection by usingtools.search, tools.get, tools.list_enabled, tools.enable, tools.disable, tools.execute, and tools.get_run. See Tools for who can manage them, how billing works, and an example call.
Share a Site
An agent can make a Site Public or Private withsites.set_visibility, so you don’t need to change its sharing in the dashboard. New Sites start Private. Activating a new code or data revision keeps the Site’s current visibility.
When a Site is Public, anyone with its link can visit it without signing in and download its complete source bundle and data. Public Sites are unlisted and marked noindex, but they are not secret or password protected. Before making a Site Public, the agent explains this and waits for you to confirm for that specific Site. The call then sends publicExposureConfirmed: true.
sites.get returns the Site’s current visibility and sharingVersion. It also returns revisionEligible, which shows whether the active revision can be made Public, and canShare, which is true only when both the MCP grant and your current permissions allow sharing that Site. Pass the current sharingVersion as expectedSharingVersion, along with a UUID requestKey. A Site without an eligible active revision can still be made Private.
If the call returns a version conflict, the change was not applied. Read the Site again before deciding whether to retry. Retrying with the same requestKey and identical input returns the original receipt, which can describe an earlier state. A Site’s canonicalUrl exists even while it is Private, so check visibility with sites.get before reporting that a Site is Public.
Apply a Template
Whentemplates.apply is available in tools/list, supply a published
templateId and an existing spaceId to create all of the template’s resources.
Sites, Automations, and ordinary File copies belong to that Space; Skills belong to the organization. Files use their configured filenames at the Space root and never overwrite existing files. A same-name conflict fails that component while independent components can continue. File completion means the bytes and metadata were durably saved; it does not wait for indexing.
Optionally supply connections as an array of { connector, connectionId }
objects, with at most one connection per provider. Supplied connections must be
accessible, compatible, and already assigned to the Space when required. Omitted
connections leave resources awaiting configuration. The response recommends
eligible connections, or creating a connection when none are available.
Creation runs in a single Trigger.dev task in dependency order, with component
outcomes recorded in the task metadata and output. Automatic retries of the
whole task are disabled; a failed or interrupted task can leave partial resources. The response acknowledges that creation
was queued and includes the planned components and connection recommendations;
it does not confirm that resources have finished creating. Use the resources’
list/get tools to find and inspect them. Successful resources remain if another
component fails.
Applying a template defers executable verification and does not activate or run
its resources. Each apply call is a new creation intent, including retries. Do not repeat
templates.apply to check progress: existing File names conflict, and executable
components may produce additional copies. An unknown write outcome interrupts the
task; inspect the destination before starting another apply. Existing successful
resources are retained, with no whole-task replay or resume binding.
templates.search, templates.get, and templates.apply include ordinary Files,
including templates containing only Files. Customer web summaries currently show
Skills, Automations, and Sites; File-only web setup is deferred. File
results appear in apply progress but cannot be executable dependency references.
File copies use the ordinary resource-file extension policy, per-file size limit,
organization storage quota, and current Space File write permissions. The task
rechecks both the delegated MCP grant and current principal permissions before
writing. A File uses the current published source at creation; later source edits
do not change existing copies. Source previews are available only to verified
platform administrators. Customers read or download their created Space files
through the existing native file tools and permissions.
Before You Connect
You need:- A Flowglad account with access to the organization you want to use.
- The Flowglad MCP URL. Use the universal URL or your organization’s URL.
- An agent or client that supports a remote MCP server over Streamable HTTP and OAuth.
- The organization and Space permissions needed for your intended work.
- Any source Connection or file already available to the relevant Space.
Choose an MCP URL
Flowglad accepts two production MCP URLs. Both expose the same organization tools after you authorize.Universal URL
organizationId field that identifies the target organization; when the grant includes multiple accessible organizations, tool calls must supply it. tools/list can also take organizationId to discover one organization’s available tools. Every operation checks both that organization’s grant selection and your current permissions.
Joining another organization does not automatically extend an existing grant. Authorize the client again to include it. If you lose access to an organization, the client loses access to it too.
Organization URL
Organization-specific URLs still work:{organizationId} with your actual Flowglad organization ID. Preserve the complete ID, including its org_ prefix. Do not enter the braces or a made-up organization ID. Flowglad skips the organization choice for this URL because the organization is already in the URL.
In Flowglad, open Settings → MCP → Add MCP Connection, select your client, and copy the MCP server URL shown there. The settings page shows the universal URL.
If you cannot access these settings, ask your Flowglad administrator or support.
A connection-specific URL can expose a narrower tool set and is not interchangeable with the universal or organization URL described here.
Manage Grants
Flowglad keeps one active universal grant for each user and client, containing all selected organizations and their individual capabilities.- Reconsent. Authorizing the same client again replaces the organization selections in its existing grant.
- Revocation. Revoking that grant ends the client’s access to every organization included in it. Existing access and refresh tokens cannot authorize MCP operations.
- Reconnect. Connecting after revocation creates a new grant.
- Under Organizations, check an organization marked Not Authorized to add it, or uncheck one to remove it.
- Expand an organization to change its permissions and capabilities.
- Select Save.
Connect from Claude Desktop
Watch the Claude Desktop setup video, or follow the steps below.
Use Claude Desktop’s custom connector flow to add Flowglad and authorize access to your account.
1. Open the custom connector form.
In Claude Desktop, navigate to Customize → Connectors → + Add → Add Custom Connector.
2. Add the Flowglad connector.
Enter Flowglad as the connector name, or choose another name.
Paste the MCP URL into the URL field. Use the universal URL,
https://dashboard.flowglad.com/api/mcp, or your organization URL from Settings > MCP, which looks like https://dashboard.flowglad.com/api/mcp/orgs/org_....
Do not change any other connector settings. Select Add to continue. Claude Desktop may label this button Continue in some versions.
3. Grant Claude access to Flowglad.
Continue through the connection flow to authorize Claude to access your Flowglad account. When prompted, sign in, choose or confirm the intended Flowglad organization, and review the requested access before authorizing it.
When the flow completes, return to Claude Desktop and confirm that the Flowglad connection is available. Ask Claude to check the Flowglad connection and list the Spaces you can access.
OAuth Dynamic Client Registration is the intended connection method. Do not add a password, access token, API key, or client secret to the server URL.
These instructions describe connecting Flowglad directly as a custom connector; directory availability is separate.
Connect from ChatGPT (Website)
Watch the ChatGPT website setup video, or follow the steps below.
1. Copy the MCP URL.
Copy the universal URL,
https://dashboard.flowglad.com/api/mcp, or open Settings → MCP → Add MCP Connection → ChatGPT and copy the MCP server URL shown there.
2. Open ChatGPT Plugins.
Go to chat.com/plugins.
3. Create an MCP app.
Select Add in the top right, then select Create MCP App.
4. Fill in the form.
Enter these values and leave the other fields as they are:
5. Create the plugin.
Check I understand and want to continue, then select Create as a plugin.
6. Sign in and connect.
Complete the Flowglad login flow. Sign in, choose or confirm the organization, and review the requested access before authorizing it.
Connect from the ChatGPT Desktop App (Codex)
The ChatGPT desktop app runs Codex locally. Codex keeps its own MCP server list, saved OAuth credentials, and tool approval settings on your computer. Flowglad does not control these settings. 1. Copy the setup prompt. While Onboarding appears in your Flowglad sidebar, open it and select Via Prompt under Connect with Your AI Agent. Copy the prompt from Copy Setup Prompt. Flowglad hides Onboarding after setup is complete. If you don’t see it, add the server yourself with the commands in the next step. 2. Paste the prompt into Codex. Paste the prompt into a Codex task in the ChatGPT desktop app. Codex adds the universal URL as an MCP server, usually namedflowglad, and starts OAuth. If you add the server yourself, run:
Find the Layer That Fails
Three different layers must all work before an agent can call a Flowglad tool:
A failure in the saved credentials or host tool permission layer does not mean that your Flowglad grant is incorrect. Do not revoke a working grant to fix a client-side error.
To recover from a saved credentials error:
- Close other Codex tasks or terminals that run a login for the same server.
- Sign in again from the MCP settings of the ChatGPT desktop app.
- If the error continues, run
codex mcp logout flowglad, then runcodex mcp login flowgladonce. - Start a new task and ask it to list your Flowglad Spaces.
- Change the Codex approval setting to a mode that asks you before tool calls. In
~/.codex/config.toml, this setting isapproval_policy. - Start a new task. The new setting does not always apply to a task that is already open.
- Ask for a read-only call first, such as a list of your Flowglad Spaces.
- For a write, ask for one small change in a test Space. Then ask the agent to read the record back.
Connect Another MCP Client
Add the universal URL or your organization URL as a remote Streamable HTTP MCP server. Complete the client’s OAuth flow with your Flowglad account, then verify the organization and available tools. If your agent can configure its own MCP servers, you can copy the setup prompt from Onboarding → Via Prompt and paste it into the agent instead. The prompt uses the universal URL and OAuth, and asks the agent to reuse an existing Flowglad connection. Client support varies. A successful server connection does not mean the client imports Flowglad Skills automatically. The Skill discovery extension and theskills.list tool are different interfaces. See Skills for supported discovery and package-management behavior.
Data Handling
Use of the Flowglad MCP is governed by the Flowglad Privacy Policy. Organization admins can see all MCP tool calls made to your organization in Settings > Audit.Troubleshooting
For general security information, see Security and our Trust Center. If you are having any trouble connecting, contact us.
When contacting support, include the time, client, affected tool, safe error code, and request ID if available. For a desktop client, also include the client name and version, your operating system, and the layer that failed: browser consent, saved credentials, or host tool permission. Never include passwords, bearer tokens, cookies, authorization codes, OAuth callback URLs, signed upload/download URLs, or customer data. Do not send a screenshot that shows the browser address bar during sign-in.