Agent access
Loan Intelligence MCP
Connect an AI agent to a purpose-built loan package engine. MCP clients can quote packages, submit source documents, poll durable status, retrieve source-linked outputs, and inspect credit state through one authenticated endpoint.
MCP connection details
MCP endpoint
Authentication uses a workspace API key in the Authorization header.
Discover
Agents and operators can start from the public MCP endpoint, discovery metadata, or machine-readable agents guide.
Credential
The workspace creates, rotates, and revokes the API key used by the MCP client.
Process
The MCP tools use the same quote, reservation, worker, billing, and result contracts as the workspace.
Runtime URLs
Agent-readable surfaces
The authenticated JSON-RPC endpoint is the processing surface. Discovery, status, and the agents guide are public runtime metadata.
MCP tools
Designed for a complete package loop
Tool names are namespaced and discoverable with `tools/list`; the workspace key page provides copy-ready client configuration after OTP access.
Quote and submit
loan_intelligence_quote_packageloan_intelligence_process_packageTrack and retrieve
loan_intelligence_get_package_statusloan_intelligence_get_package_resultsloan_intelligence_list_packagesCredits and policy
loan_intelligence_get_credit_balanceloan_intelligence_get_credit_policyloan_intelligence_update_credit_policyFAQ
MCP setup questions for operators.
An agent-ready MCP surface needs more than a URL. These answers cover the credential, header, tools, document payload shape, polling loop, credit behavior, and the internal details clients should never see.
Use the JSON-RPC endpoint at /api/plugins/mcp/mcp on the Loan Intelligence origin. The public /mcp page shows the live endpoint, discovery metadata, status route, and agents guide for the current environment.
Create or sign into a workspace account, open API & MCP keys, create a Stack API key, and copy it immediately. The raw secret is shown once. The same page provides a copy-ready mcpServers.loan-intelligence configuration for clients that accept remote MCP servers.
Send Authorization: Bearer <your-api-key>. Public discovery and status metadata can be read without a key, but package processing, result retrieval, credit balance, and credit-policy tools require an authenticated workspace user.
The current Loan Intelligence namespace exposes tools to quote a package, process a package, get package status, get package results, list packages, get credit balance, get credit policy, and update credit policy. The status route reports the live tool count so clients can verify the contract before running.
Yes. The process tool accepts sourceDocuments with fileName, optional page counts and metadata, and inline content as contentBase64 or contentText. Real legacy-engine processing currently requires staged source payloads for each source document, so metadata-only submissions can quote but are not enough for the worker path.
Store the returned packageId and session state, poll get_package_status until the package is completed or failed, then call get_package_results for result manifests, source references, facts, warnings, tasks, and output links. Agents should surface failures directly instead of retrying blindly when policy, credits, or worker dependencies fail closed.
MCP responses are shaped for operators and agents. They return public package, status, result, source-reference, and credit fields, but they do not expose raw queue keys, cache heads, staged payload text, AppData locations, internal idempotency keys, or worker implementation details.
The agent should quote first. Processing reserves credits before work starts and captures them only after durable billable outputs exist. The get_credit_balance tool returns the authenticated user's Stack billing balance, and workspace credit policy can require quote confirmation or cap the maximum package reservation.