Integrate.io MCP Server¶

Data work breaks in the gap between what a pipeline does and who can read it. The engineer who built it knows why the Salesforce sync failed at 04:00. The founder asking the question gets a dashboard that is three hours stale and nobody free to explain it.

Integrate.io puts its ETL and Reverse ETL platform behind one hosted MCP endpoint, so the question about the pipeline is answered in the same conversation as the question about the business. The endpoint is per account and the URL tells you why:

Server type: Remote (Streamable HTTP)
Endpoint:    https://api.xplenty.com/YOUR_ACCOUNT_ID/api/mcp
Transport:   one JSON-RPC 2.0 message per request, batching not supported
Auth:        OAuth 2.1 browser sign-in, or a Personal Access Token
Tools:       46 per the vendor documentation (44 on the directory card)
Clients:     Claude, Claude Code, ChatGPT, Cursor, Perplexity, Codex, any MCP client
Billing:     included with an Integrate.io ETL account
Site:        integrate.io/platform/mcp/

The domain is api.xplenty.com because Integrate.io was formerly Xplenty and the API predates the rebrand. The dashboard URL and the API URL both work, and the account id can be copied from the dashboard rather than assembled by hand.

A keyless initialize against an unpopulated account path answered HTTP 401 with Bearer error="invalid_token" and resource_metadata="https://api.xplenty.com/0/.well-known/oauth-protected-resource", which is a live OAuth-gated endpoint naming its own metadata rather than an empty refusal. No tool was called: the surface needs an account, so this entry is documented from the vendor's parameter-level reference and states that tier plainly.

The tools split three ways, and the split is the design¶

Reading and inspecting, safe on a production account: list_connections and get_connection for the shape of a source without its credentials, list_workspaces, list_packages and get_package for the flow graph with secrets redacted, list_clusters, search_all for free-text search across connections, packages, clusters and workspaces, validate_package and validate_expression, discover_schema and discover_file_schema, preview_data, preview_transformation, profile_data, test_connection, ask_docs for a sourced answer out of the live documentation, and report_issue carrying the call_id the server attaches to every result.

Building and editing, which is where a description becomes a pipeline: create or clone a package, add, edit and remove components and edges, manage variables, save or discard a draft, and roll back a version.

Jobs and schedules, where the operational questions live: list_runs, get_run, get_run_summary, explain_job_failure, list_schedules and get_schedule, then toggle_schedule, create_schedule, update_schedule, create_cluster, run_package, cancel_job and retry_job.

Two of those deserve their own line. get_run_summary returns records read per source and written per destination, per flow, so a partial load is visible instead of inferred. explain_job_failure returns the primary error, the detected error patterns and the log tail, which is the difference between an assistant saying a job failed and an assistant saying why.

The containment details are what an operator should read first¶

The vendor is unusually direct: mutating tools take effect immediately and there is no sandbox. A package edit or a schedule change is against production. Three published behaviours make that survivable rather than alarming:

Tools that subclass the mutation base support a dry_run, which validates and returns the would-be outcome without applying it. Any agent workflow that touches a live pipeline should set it first, and the guide says so because the default is the opposite.
create_schedule creates a schedule disabled by default, so a schedule built by an assistant does not start firing because the tool succeeded.
create_cluster provisions billed infrastructure. It is a spend action wearing a provisioning name, and it belongs behind the same approval a payment would get.
Package edits are versioned and can be rolled back, and every change made through MCP lands in the account's audit history.

Two smaller traps are documented rather than discovered later: list_runs defaults to root runs only, so child retries are excluded unless you ask for them, and validate_package records a validation and writes the workflow's script to S3 even though it does not alter a pipeline.

Access follows the person, not the token: every tool is scoped to the authenticated account, obeys the signed-in user's permissions, and a token for another account gets a 403. The tools return connection shapes but never credentials, TLS certificates or SSL keys.

On the token lifetime, and why it matters more than it looks¶

The vendor supports OAuth 2.1, which most modern clients use, and a Personal Access Token for everything else. The PAT default lifetime is ten years. For an account that can build pipelines and start billed clusters, a decade-long static credential is the weakest link in an otherwise careful permission model, and it is the one setting a buyer should change on day one. Revocation and re-authentication are documented on the same page, which is the right place for them.

Why this earns a guide¶

There is no ETL entry in the catalogue. The neighbours read data rather than move it: SaturnSQL MCP and Snowflake MCP query a warehouse that something else has to fill, and SQL Server MCP runs a DBA console over instances.

The pipeline that fills the warehouse was the uncovered angle, and it is the one a business operator hits as "the numbers in my assistant are from yesterday". A connector can read a source directly; only the pipeline knows what was transformed, what loaded and what failed between them. Reading that from chat, and repairing it from chat with a dry run, is the operator surface nothing else in these 1,094 guides covers.

FAQ¶

Does the server need to be installed or hosted?¶

No. Integrate.io hosts it. You add your account's URL to the client and sign in, and stdio-only clients reach it through an mcp-remote bridge.

Can an assistant change a live pipeline?¶

Yes, and it takes effect immediately because there is no sandbox. Mutation tools accept dry_run, schedules are created disabled, and every change is versioned and audited.

Which tool counts are right, 44 or 46?¶

The vendor's documentation says 46 tools across the three areas; the directory card says 44. The vendor also tells you to run tools/list against your own account for the exact set, which is the answer that cannot drift.

What credential should I use?¶

OAuth 2.1 where the client supports it, since nothing is pasted or stored. If a static token is required, mint a PAT and shorten its ten-year default before using it on an account that can start billed clusters.

Integration with CorpusIQ¶

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects