9. InferrexMCP

InferrexMCP is a Model Context Protocol server. It gives an AI coding tool the same operations as InferrexAPI, as tools. Each tool is one API operation, runs as your workspace, and is held to the same permissions.

The server is at:

https://mcp.inferrex.com/mcp

It speaks Streamable HTTP.

Connect an AI tool

There are two ways to authenticate. Choose one.

In your AI tool, add an MCP server with the address https://mcp.inferrex.com/mcp.
The tool opens the Inferrex sign-in. Sign in.
On "Connect an assistant", check the workspace and the permissions asked for, then approve.

There is no key to copy. Your workspace role still limits every call the assistant makes.

To see or disconnect an assistant: Workspace settings, Connected assistants.

API key (for a machine)

Send an API key in a request header. Either of these works:

X-API-Key: <your API key>
Authorization: Bearer <your API key>

How you set a header depends on your AI tool. Keep the key in the tool's secret store or an environment variable.

A key in the address is refused. https://mcp.inferrex.com/mcp?apiKey=... does not work, because keys in addresses end up in logs and browser history.

A session is tied to the key that opened it. If the key is revoked, the session ends and the next call answers 401.

Clients that only speak the older SSE transport can connect to https://mcp.inferrex.com/sse with a key in a header. Signing in works on /mcp only.

How the tools are named

Tool names are snake_case and start with their area:

Area	Tool names start with	Examples
Systems	`connection_`, `connections_`	`connection_infer`, `connections_test_connection`
Credentials	`credential_`, `credentials_`	`credential_store`, `credentials_list_failing_secrets`
Mappings	`mappings_`	`mappings_get_mapping`, `mappings_approve_mapping`
Flows	`pipeline_`, `pipelines_`	`pipeline_create`, `pipeline_dry_run`
Golden records	`unified_api_`, `golden_record_`	`unified_api_unified_get_action`, `golden_record_approve_proposal`
Repairs	`heals_`	`heals_list`, `heals_answer_heal`
Health and safety stops	`monitoring_`	`monitoring_get_connection_health`
Jobs	`jobs_`	`jobs_get_job`
Workspace	`workspace_`	`workspace_get_environment_overview`

Most tools take the record's id as id. Each tool's own description says what it does, what it needs and what it returns. Each tool is also marked as read-only or not, and as destructive or not. A well-behaved AI tool asks you before it runs one that is not read-only.

When a call fails, the tool returns the API's error body (chapter 8) and marks the result as an error.

A first session

Ask your AI tool for an overview. It should call:

workspace_get_active_environment to confirm which environment it is in.
workspace_get_environment_overview with detail: "brief" for what needs attention.

Then work through the sequence. These are the tools for each step.

Connect a system

Step	Tool
Ask the API how it signs in	`connections_detect_auth`
Test the address and sign-in. Nothing is stored	`connections_test_connection`
Connect and start the first read	`connection_infer`
Wait for the read	`connection_status`, `jobs_get_job`
Read what was found	`connection_manifest`
Accept or correct a field	`connection_approve_field`
Leave fields out	`connections_set_field_inclusion`
Approve the system	`connection_approve`

Map and build a flow

Step	Tool
Match two systems	`connections_match`
Read the mapping	`mappings_get_mapping`, `mappings_list_mappable_fields`
Decide links	`mappings_update_field_mapping`, `mappings_bulk_update_field`
Approve the mapping	`mappings_approve_mapping`
Create a draft flow	`pipeline_create`
Dry run. Nothing is written	`pipeline_dry_run`, then `jobs_get_job`
Turn it on	`pipeline_activate`
Sync now	`pipeline_sync`
Set when it syncs	`pipelines_set_pipeline_schedule`

Review and watch

Step	Tool
List Golden-record changes waiting	`golden_record_canonical_change_proposals_list`
Read one change	`golden_record_canonical_change_proposals_get`
Decide it	`golden_record_approve_proposal`, `golden_record_reject_proposal`
Trace a value to its source	`golden_record_get_golden_field_lineage`
Health of each system	`monitoring_get_connection_health`
Flows with something waiting	`pipelines_get_pipeline_attention`
A flow's runs	`pipeline_history`
Failed writes	`pipelines_list_pipeline_quarantine`, `pipelines_list_quarantine`
Repairs and their questions	`heals_list`, `heals_get_heal`, `heals_answer_heal`
Safety stops	`monitoring_stream_circuit_breakers_list`, `monitoring_list_circuit_breaker_records`

Credentials and the conversation

An AI tool sees everything you type and everything a tool returns. You choose how a credential reaches Inferrex.

Save it as a secret first. This is the safest way. Save the credential in the app (Workspace settings, Secrets), then give the assistant the secret's name. connections_test_connection and connection_infer take that name as credentialRef, and the value never passes through the assistant.
Paste it into the conversation. This also works: the assistant passes the value on in credentials. Know what it means before you do it. The AI tool has then seen the value, and the conversation's history may keep it. The Inferrex skills warn you and ask you to confirm before they accept a pasted credential.
Inferrex never returns a stored credential. Tools that list credentials return names and dates, never values.
A new API key or signing secret is shown once, in the tool's answer. Create those in the app when you can.

What needs a person

Some steps cannot be finished by an assistant alone:

A provider's own sign-in page. A person must sign in and approve in a browser. Start it from the system's Sign-in tab in the app.
Approvals. Where a policy names approvers, the person who asked for a change cannot approve it. An assistant acting as you has the same limit.
Saving a secret's value without the assistant seeing it. Do it in the app, as above.

Use skills with InferrexMCP

The skills page holds ready-made instructions for AI coding tools: connect a system, map two systems, build a flow, review Golden-record changes, diagnose a failing flow and replace credentials. Each one names the tools to call, in order, with the checks to make before any write.

Next: InferrexCLI, the cloud shell.