Deploy a Model Context Protocol endpoint with Instant's one-click tools
Connecting an AI client to a site-building workflow often means switching between prompts, dashboards, and manual publishing steps. An Instant MCP endpoint reduces that friction by letting compatible AI tools connect to Instant’s hosted Model Context Protocol server and work with the landing-page capabilities Instant exposes.
The important distinction is that this is not a generic, unrestricted storefront automation layer. Instant’s MCP support is in beta and currently focuses on landing pages, so the right deployment plan is to connect the hosted endpoint, authorize it carefully, verify its available actions, and keep a human review step before publishing customer-facing changes.
What an Instant MCP endpoint deploys and why it matters
Model Context Protocol (MCP) is a standard for exposing tools that language models can invoke. In the MCP specification, servers expose tools that models can call through the tools/call endpoint. In practical terms, an MCP server gives an AI client a defined way to discover and use actions provided by another system.
Instant provides a hosted remote MCP server. Rather than building and hosting an MCP service yourself, you add Instant’s documented endpoint to a supported client and authorize access. Instant describes this hosted remote server as the easiest way to get started, with OAuth used to grant access to an Instant account.
Direct answer: To deploy an Instant MCP endpoint connection, add Instant’s hosted URL,
https://mcp.instantdb.com/mcp, to a compatible AI client, complete the authorization flow, and then use the client to access the Instant tools available for your project. For landing-page workflows, project access tokens are the documented credential mechanism for external AI clients and are created in Instant under Settings → General → Access tokens.
The phrase “one-click” is best understood as low-friction setup, not as a promise that every publishing task requires no decisions. The hosted server removes the need to operate your own remote MCP infrastructure. You still need to select a client, establish access, understand the scope of the tools, manage credentials, and check generated page changes.
For teams already using AI coding or agent tools, that difference is meaningful. An AI workflow can use Instant capabilities to create landing pages, update layouts, generate sections, and change storefront content where those actions fall within the currently supported landing-page workflow. The MCP connection puts those capabilities closer to the place where prompts and implementation decisions are already happening.
Know the current Instant MCP scope before connecting
Start with the boundary, because it determines whether MCP is the right route for the task in front of you. Instant explicitly states that MCP currently works with landing pages only.
The documented MCP tools can perform several useful landing-page operations:
List landing pages.
Create landing pages.
Edit landing pages.
Publish and unpublish landing pages.
Send natural-language instructions to Instant’s AI page-builder agent.
Fetch page details, project information, and the project’s AI credit balance.
That set supports a focused workflow: inspect a landing page, ask for a change, evaluate the result, and decide whether to publish. It can also support creating a campaign-specific page when a team has clear copy, layout, and brand instructions ready to give the AI client.
What is outside the current scope
Do not plan an Instant MCP deployment as if it controls every part of a storefront. The current support does not cover product or collection templates, sections, ers and footers, the cart, or the storefront home page.
This limitation is not a minor implementation detail. If a campaign depends on changes to navigation, a product-template structure, collection presentation, cart behavior, or the home page, those changes need another workflow. MCP can still be useful for a dedicated landing page, but it should not be presented internally as a replacement for all storefront administration.
A practical way to choose is to ask a narrow question: “Can this deliverable live entirely on a landing page?” If the answer is yes, Instant MCP may be an efficient fit. If the work crosses into the excluded storefront areas, use the relevant Instant interface or another approved process for those parts.
Choose hosted remote MCP or a local Instant MCP server
Instant documents two ways to connect. The best choice depends less on preference for a particular protocol and more on where the AI client runs, how your organization handles authentication, and whether you want the simplest managed starting point or a local credential-based setup.
Hosted remote MCP: the simplest documented starting path
Instant recommends its hosted remote MCP server as the easiest way to start. The streamable HTTP endpoint is:
https://mcp.instantdb.com/mcp
Instant also documents an SSE endpoint:
https://mcp.instantdb.com/sse
Adding the hosted remote server routes the user through OAuth to grant access to the Instant account. This is the natural choice when your client supports the remote endpoint and OAuth flow, and you do not need to run an MCP package locally.
Local stdio: an alternative when local execution is required
Instant also supports running @instantdb/mcp locally over stdio. Instant says this path avoids OAuth, but it requires a personal access token. That trade-off matters: local execution may fit a client or environment that expects stdio, while it shifts credential handling to the user or team operating the local process.
For SSE-based setups, Instant provides examples using a local mcp-remote option in addition to its hosted endpoint examples. Windsurf and Zed are among the clients for which Instant publishes configuration guidance.
Make the decision deliberately
Use the hosted remote endpoint when you want Instant’s documented easy-start path and your AI client supports remote MCP authorization.
Use the documented SSE endpoint only when your client or setup calls for SSE transport.
Consider local
@instantdb/mcp
via stdio when local execution is a requirement and your team can protect and manage a personal access token.
Do not select a path solely because it sounds more automated; select one that matches your client’s documented MCP transport and your access-control process.
In all cases, the endpoint is the connection point, not a substitute for governance. A working connection should be followed by a small, reversible test before anyone asks an AI agent to make broad content changes.
Deploy the hosted Instant MCP endpoint in your AI client
The exact interface differs by client, but the deployment pattern is consistent: add the Instant server, use the documented URL and transport, authorize it, then confirm that the client can see and use the available tools. Start with one project and a low-risk landing page rather than connecting a production workflow and immediately publishing changes.
Confirm the intended project and task.
Identify the Instant project and make sure the work is limited to landing pages. Decide what “done” means, such as creating an unpublished page or revising existing copy.
Select a compatible client.
Instant documents setup paths for Cursor, Claude Code, Codex, Gemini, Windsurf, and Zed. Use the instructions appropriate to the tool your team already works in.
Add the hosted remote server.
For streamable HTTP configurations, use
https://mcp.instantdb.com/mcp
. Use the documented SSE URL only in a setup that expects it.
Complete authorization.
With the hosted remote option, Instant says the connection routes through OAuth to grant access to the Instant account. Carefully verify the account and project context before approving access.
Validate with a read-oriented action.
Fetch project information, page details, or the AI credit balance. These checks let you confirm the connection before asking an agent to alter a page.
Run a constrained content test.
Ask for a specific draft-level landing-page change, then inspect the result. Publish only after the output meets your requirements.
Instant’s published examples make the remote configuration concrete. For Cursor, Instant shows a configuration that points MCP to https://mcp.instantdb.com/mcp. The details of where you paste configuration depend on the Cursor environment, but the endpoint itself is the key shared value.
For Claude Code, Instant documents this command:
claude mcp add instant -s user -t http https://mcp.instantdb.com/mcp
For Codex, Instant documents enabling the rmcp_client feature and then adding the server with:
codex mcp add instant --url "https://mcp.instantdb.com/mcp"
For Gemini, Instant documents:
gemini mcp add --transport http instant https://mcp.instantdb.com/mcp
These commands and configurations should be treated as client-specific starting points, not interchangeable syntax. Follow the setup requirements in the client you are actually using, including the Codex feature requirement where applicable. If a tool cannot connect with the transport you selected, check the client’s MCP support and use one of Instant’s documented alternatives rather than guessing at configuration values.
Create and protect Instant access tokens for MCP workflows
Access tokens are central to Instant’s documented connection model for external AI clients working with an Instant project over MCP. This capability is labeled beta, which is a good reason to use a deliberately narrow rollout: involve only the people who need access, test on a limited workflow, and document how access will be removed.
Only project owners and admins can create or revoke access tokens. The documented dashboard path is straightforward:
In Instant, open
Settings → General
.
Open
Access tokens
.
Click
New token
.
Name the token so its purpose is recognizable later.
Copy the full token immediately and store it in an approved secure location.
Instant shows the full token only once. That makes the moment of creation an operational handoff: the person creating it should know where the token belongs before clicking the button. Avoid placing a token in a prompt, a source-controlled configuration file, a chat transcript, or other locations that are not intended for secrets.
Understand who and what a token represents
An Instant access token is scoped to the current project and tied to the member who created it. If that member loses access to the project, Instant automatically revokes that member’s tokens for the project. This behavior helps align token access with project membership, but it also means that offboarding or role changes can affect an active workflow.
Name tokens for accountability and maintenance, not convenience. A name such as “marketing-assistant-testing” communicates more than a generic label and makes it easier for an owner or admin to identify whether the credential still has a valid purpose.
Plan revocation before you need it
Instant says token revocation removes access immediately and cannot be undone. That is exactly what you want when a credential is no longer required or may have been exposed, but it means revocation should be coordinated with anyone who depends on the connection.
Create tokens only for a defined project and workflow.
Keep the token with the minimum necessary group or system.
Record the token’s owner, purpose, and the client or local process that uses it.
Revoke a token when the workflow ends, the creator’s role changes, or you suspect disclosure.
Expect a replacement token and reconnection process after revocation; access does not return by undoing the revocation.
The remote OAuth flow and local personal-access-token workflow are documented as distinct connection approaches. Whichever method you use, do not blur authentication convenience with broad authorization. Confirm that the account, project, and client session are all the intended ones before making changes.
Manage AI credits and publishing risk in landing-page automation
An MCP connection can retrieve information and can also invoke Instant’s AI page-builder agent. Those two categories have different operational consequences. Instant states that tools which run the AI agent draw from the project’s AI credits, while reading data and checking status do not consume credits.
Before asking an agent to generate or revise a landing page, check the project’s AI credit balance. This is one of the data points the MCP tools can fetch, and it helps prevent a workflow from failing midway through a planned task because credits are unavailable.
If a project is out of credits, the MCP agent will not run until credits reset or more are purchased. Instant also notes that the agent will not run on projects with no credits, including the Free plan. A successful endpoint connection therefore does not guarantee that an AI-agent request can execute.
Separate inspection, generation, and publication
A sound workflow treats these as three separate stages, even though MCP can make them feel adjacent:
Inspect:
retrieve page details, project information, and credit status. Clarify the page, audience, required content, and constraints.
Generate or edit:
send narrow natural-language instructions to the AI page-builder agent or use the available page operations. This is the credit-consuming stage when the AI agent runs.
Review and publish:
inspect the output for accuracy, brand fit, legal or policy requirements, links, calls to action, and visual intent before using the publish action.
This sequence is especially important because Instant MCP can publish and unpublish landing pages. The ability to publish is useful for reducing handoffs, but it is also the point where an internal draft becomes customer-facing. A human approval step is a practical control, not an obstacle to automation.
Prompts should be specific enough to make review easier. Instead of asking an agent to “improve the page,” identify the requested changes: the target audience, the value proposition to emphasize, content that must remain unchanged, the desired call to action, and whether the expected result should remain unpublished. Specific instructions do not eliminate review, but they give reviewers a clearer standard against which to judge the result.
Use Instant MCP with rules and a repeatable team workflow
Instant recommends combining MCP with its rules file to build full-stack apps directly in an editor. That recommendation points to a broader implementation principle: the MCP endpoint supplies tool access, while rules and team conventions provide the context that makes an AI workflow more consistent.
For landing-page work, the rules file can sit alongside a lightweight operating process. The goal is not to add ceremony to every edit; it is to give the AI client and the human operator clear boundaries before an instruction turns into a published page.
A practical operating pattern
Define the request:
identify the landing page, campaign goal, approved source material, and publishing expectation.
Check the boundaries:
confirm the work does not require product templates, collection templates, ers, footers, the cart, or the storefront home page.
Verify access and status:
make sure the correct project is connected and inspect AI credit availability before invoking the agent.
Use constrained instructions:
state what should change and what must not change. Keep the first request small enough to review quickly.
Review the page:
validate copy, content accuracy, links, formatting, brand requirements, and campaign requirements.
Publish intentionally:
publish when approved, or leave the page unpublished while changes are still being evaluated.
Remove access when appropriate:
revoke tokens that are no longer needed, remembering that revocation takes effect immediately.
This pattern also clarifies responsibility. The AI client can invoke tools, but a team still needs someone accountable for the request, the credential, the review decision, and the final publish action. That is particularly relevant for beta functionality and customer-facing content.
MCP adoption is receiving attention across the ecosystem: Google documents MCP in the context of server proxies for external services, while AWS guidance and recent research discuss automation and enterprise deployment patterns. Those broader discussions do not change Instant’s product boundaries. They do, however, reinforce the value of treating MCP connections as integrations with identity, authorization, change-control, and observability considerations rather than as simple chat extensions.
Troubleshoot an Instant MCP connection without expanding scope
Most early issues fall into a few categories: the client is not configured for the selected transport, authorization has not completed, the user lacks the necessary project role, the wrong project context is active, the requested task is outside landing pages, or the AI agent has no available credits.
Work through the simplest checks first:
Confirm the endpoint exactly matches the documented URL:
https://mcp.instantdb.com/mcp
for streamable HTTP, or
https://mcp.instantdb.com/sse
for the documented SSE endpoint.
Confirm that the selected client and command or configuration match Instant’s published setup guidance for that client.
For a hosted remote connection, complete the OAuth authorization and verify you used the intended Instant account.
For token-based use, check that an owner or admin created the project’s access token, that it was copied at creation time, and that it has not been revoked.
Test a non-credit-consuming status or data read, such as project information, page details, or credit balance.
Check whether the requested action belongs to a landing page. If it concerns a er, footer, cart, home page, product template, collection template, or those excluded areas, MCP is not currently the right tool for that task.
When invoking the AI agent, check the AI credit balance. The agent cannot run when the project has no credits.
If a team member unexpectedly loses access, remember the creator-tied behavior of project tokens: when the token creator loses project access, Instant automatically revokes that member’s tokens for that project. In that situation, investigate the membership change and establish a replacement credential through an authorized owner or admin if the workflow still needs one.
Finally, avoid using troubleshooting as a reason to broaden permissions or move a secret into an insecure place. A connection problem is best solved by checking endpoint, transport, authorization, project context, scope, and credits in that order,not by distributing credentials more widely.
Instant’s one-click-oriented MCP tooling is most useful when the job is clearly a landing-page job: connect a supported AI client to the hosted endpoint, authorize access, inspect the available page context, and use the agent or page tools within a reviewed publishing process. The hosted remote server offers the simplest documented starting point, while local stdio provides an alternative for environments that need local execution.
Begin with a single low-risk page, confirm your client can read project status, verify available AI credits, and keep the first generated change unpublished until it has been reviewed. From there, use clear rules, purposeful token management, and the current landing-page-only scope to make the Instant MCP workflow useful without assuming it controls the entire storefront.