Skip to content
CapitalSourceDeveloper platform
MCPProtocol version 2025-06-18 · stdio + Streamable HTTP · 14 tools

The protocol layer
for agentic credit.

Any agent that speaks the Model Context Protocol can run a CSG-1003 intake without writing a single REST call. Point your own platform at the HTTP transport with an OAuth2 client you mint and rotate yourself, or attach a client on your machine over stdio. Same tools, same ledger, same audit trail either way.

How your stack connects See the 14 tools
Two ways in

Bring your agent stack.

The tools are the same either way. What differs is who is calling them — a platform you run, or a client on your desk.

Your platform calls the tools

Brokers, lenders, and anyone running their own agents

planned

Built and tested. Not reachable yet — no public hostname is provisioned.

A Streamable HTTP transport at POST /mcp, stateless, so nothing has to outlive a single request. Your service authenticates with OAuth2 client credentials — a client you mint and rotate yourself, not a personal token pasted into a config file. Discovery at /.well-known/mcp.json tells your agent the endpoint, the protocol version, the auth flows and where to verify our signatures.

  • OAuth2 client credentials, minted and rotated from the portal
  • Stateless: one bearer token per request, no session to keep alive
  • Discovery document carries the endpoint, auth flows and trust JWKS
  • Every call lands in the same usage ledger and audit log as a REST call

A client on your machine

Trying it out, or working a deal by hand

live

Works today, over stdio, with any client that can launch a subprocess.

The Recurser CLI doubles as an MCP server over stdio. Point any MCP-speaking client at the binary and the 14 tools register. Claude Desktop is the one client we write the config block for automatically; every other client you point yourself.

  • Any stdio client works — you point it at the binary
  • Claude Desktop has a config writer: recurser mcp install
  • Same bearer token as REST, no separate credential lifecycle
  • The CLI is private at 0.0.0 and unpublished, so you build it from the repo

CapitalSource Agent Harness · Processor pilot

Connect your agent stack. Understand the execution layer.

MCP exposes CapitalSource application and capability tools to your own agent harness. For the Processor readiness pilot, the CapitalSource Agent Harness runs in Core and manages saved progress, approval waits, and execution limits.

Explore the agent harnessSee the integration surface

Where it fits your stack

  • Use the registered MCP tools for supported CSG-1003 application and capability operations.
  • Use REST, the TypeScript SDK, or the portal to inspect and control Processor harness tasks.
  • Choose the interface for each operation while Core evaluates the caller’s permissions.

Harness task controls are not currently registered MCP tools. Calling a capability through MCP does not automatically create a harness task.

What the server exposes

Tools today. Resources and prompts next.

MCP is built around tools (what an agent can do), resources (what an agent can read), and prompts (canonical instructions a host can offer). We register tools. The other two primitives are zero — here is the honest count.

14

Tools

live

Lifecycle verbs plus capability calls, same dispatch path as REST.

0

Resources

planned

None registered. Read state via apps_get and apps_package.

0

Prompts

planned

None registered. The intake prompt stays server-side.

Tool catalog

Fourteen tools. Nine of them functional.

Seven application-lifecycle tools plus seven capability tools. Five capability tools are contract-locked stubs: schemas frozen, billed at zero, returning status: "pending" so you can build against the shape before the engine lands.

Application lifecycle · 7 tools
  • pinglive
    Calls the Core Service /v1/ping endpoint and echoes the timestamp.
  • apps_createlive
    Create a draft CSG-1003 application; returns its id + handle.
  • apps_getlive
    Snapshot: phase, filled and missing fields, qualifying products.
  • apps_chatlive
    Message the Intake agent; SSE stream collected as typed events.
  • apps_qualifylive
    Run the gating engine; flips status to qualified.
  • apps_certifylive
    Sign and persist a package. FCRA authorization fields required.
  • apps_packagelive
    Latest package: canonical JSON, signature, signed storage URLs.
Capabilities · 7 tools · 2 functional
  • score_csg_1003_completenesslive
    Completeness + gating score. No LLM call.
  • generate_credit_memolive
    Memo with every figure cited to a CSG-1003 field id.
  • analyze_bank_statementsstub
    Cashflow summary from statement PDFs and CSVs.
  • extract_financials_from_documentsstub
    Financials from returns, P&Ls, AR aging.
  • verify_business_identitystub
    KYB existence, standing, OFAC screening.
  • detect_statement_fraudstub
    Forgery indicators across statement files.
  • calculate_factoring_borrowing_basestub
    Lendable amount from an AR-aging report.

Tools forward to the same REST endpoints the Recurser CLI calls, so there is one dispatch path rather than two implementations. The three-surface parity workflow reports on every PR but skips until the parity suite under apps/core/test/parity/ is written — byte-identical envelopes are the design, not yet a live assertion. There is no agent.run tool either: the eleven built catalog agents do not all have MCP bindings. Only Intake is reachable through this transport, using apps_chat.

Local path

Running it on your machine.

The Recurser CLI doubles as an MCP server. recurser mcp serve speaks stdio, so any client that can launch a subprocess can register the tools — point yours at the binary and it works. For Claude Desktop specifically, recurser mcp install writes the config block for you, carrying the token you logged in with.

  • The config writer covers Claude Desktop today. Every other stdio client works; you point it at the binary yourself.
  • recurser mcp serve is stdio only; the HTTP transport is a separate service.
  • The CLI is still private at 0.0.0 and not published, so today you build it from the repo. See how to build it.
  • Same PAT auth as REST — no separate credential lifecycle.
  • Tools forward through the REST surface, so a tool call lands in the same usage_events ledger as a REST call.
~ recurser mcp
# Serve the tools over stdio — any client can attach
recurser mcp serve
recurser mcp serve: connected to $CSG_CORE_URL via stdio
# Claude Desktop has a config writer
recurser mcp install --client claude-desktop
recurser mcp install: added entry at ~/.config/Claude/claude_desktop_config.json
Get connected

Your agent stack. Our credit protocol.

MCP is how the credit ecosystem becomes agentic — every actor (broker, funder, underwriter, servicer) talking to every other actor through the same canonical schema. CSG-1003 is the schema. The MCP server is the bridge.

Install Recurser + MCP Read the CSG-1003 protocol