Ask Claude

Ask your portfolio anything.

The first MCP-native intelligence layer over fund-compliance + portfolio-monitoring data, built for emerging-market and DFI lenders. Open Claude. Add the connector. Ask a question. Every figure cites the row it came from, matches your dashboard, and stays inside read-only permissions.

The number you want is already in the system. Getting it out has always meant knowing where to look.

Compliance Intelligence MCP, live in production

Grounded, not generative guesswork.

One question, the tool Claude called to answer it, and the row every figure came from. This is the shape of every answer, not a highlight.

claude desktop / capitalbridge connector Connected
Youcredit officer
What's deteriorating across the book?
Claudereading portfolio_compliance
Three borrowers entered the watchlist this week. Northern Energy row 47 DSCR 1.22x against a 1.20x minimum, 2% headroom remaining. Borealis Health row 123 LTV 71% against a 75% cap, 6% headroom. Apex Capital row 89 ICR 1.46x against a 1.50x minimum (breached) on the Q3 QMA submission, dated 2026-06-12. The credit committee chair has been notified by the alert engine.

Source: portfolio_compliance.borrowers, filtered to at-risk and breached, 3 rows, read-only. Latest data 15 June 2026, 09:23 UTC. The same figures the PM Portal shows.

The buying objection

No hallucinated numbers. Ever.

Buyer guides warn that general legal AI "misstates covenant levels with confidence." The CapitalBridge MCP is built so that cannot happen. Three structural choices:

01
Tools, not generation
The MCP exposes specific read tools (borrower_profile, covenant_health, portfolio_compliance) that return structured rows from the same AppServices that power your dashboard. The model selects which tool to call; it does not invent the numbers.
02
Citation required
Every figure returned by a tool carries a row identifier and a timestamp. They are surfaced in the assistant's answer as clickable references (Northern Energy), so a credit officer can verify each number against the source in one click. Unciteable claims do not appear.
03
Dashboard parity
The MCP and the web portal read from the same query path through the same AppServices, with the same tenant + scope filters. A number returned in chat is identical to the number a credit committee member sees in the portal at the same moment. No drift. No reconciliation work.
Try it

Four canned questions, four grounded answers.

These are real questions production users ask, with sample answers from anonymized portfolio data. Click any question to see the response, the rows it cites, and the provenance block.

Try a question

Anonymised sample data. Read-only. No personal data.

How it works

Put your compliance book in Claude.

One governed connector over the book you already run. Your team signs in as themselves, the assistant sees only what each role allows, and every call is written to the log whether it was answered or refused.

  1. 01Connect your book

    One MCP server over the funds and facilities you already run.

  2. 02Ask in plain language

    Live answers from your own data, every figure traced to a row.

  3. 03Approve every change

    Writes are drafted and held until a person signs off.

  4. 04Delivered in your brand

    Factsheets, board packs and workbooks on your own template.

  5. 05Only what you allow

    Named actions per server, never free-form database access.

Works with Claude, Microsoft Copilot, and ChatGPT.

01

Connect your book

One governed connector over the book you already run. Your team signs in as themselves; the assistant sees only the funds each role already opens in the portal.

02

Ask in plain language

No SQL, no exports, no new tool to learn. Every figure traces to a returned row, and where a value is missing the answer says so rather than filling the gap.

03

Approve every change

Writes are drafted and held, never fired. You preview the diff, you sign off, it commits, and the call is written to the log with the document page it was read from.

04

Delivered in your brand

Quarterly factsheets, board packs, covenant workbooks and charts come back on your own template, in your colours, straight from the assistant. The renderer is deterministic code, so the same quarter always draws the same page.

05

Only what you allow

The actual permissions screen in Claude: every named tool is allow, ask first, or off, per user, and the list is everything the assistant can do. In production today CapitalBridge runs 88 named tools: 74 read-only and 14 guarded writes.

The wedge

The incumbents are developed-market, no-MCP, no-impact.

CovenantIQ, Allvue Nexius / Document IQ, and Octus CovenantAI are excellent for developed markets, with clean APIs to fund-accounting backbones and no need for impact-covenant logic. Cardo AI is AI-native portfolio analytics for private debt and securitisation. None of them ship an MCP-native conversational delivery layer, and none of them solve for the EM/DFI buyer.

CapitalBridge is the only platform combining EM/DFI focus, impact covenants (HIPSO / JII / SDG adjacent), facility-agreement variety, and MCP-native conversational delivery. The white-space line: DFIs are unserved at the operational layer because the industry energy is at the macro/policy level, not portfolio ops.

Common Questions

Questions credit teams ask

Can't find yours? Book a call and ask it directly.

What does MCP-native conversational compliance intelligence mean?+

MCP (Model Context Protocol) is an open protocol that lets AI assistants like Claude securely read from external data sources. CapitalBridge ships an MCP server that exposes read-only fund-compliance and portfolio-monitoring data: borrower compliance, covenant health, portfolio-wide deterioration. A credit officer can ask Claude "what's deteriorating across the book?" and get a grounded answer with row-level citations. Every figure traces to a returned row, nothing is invented by the model. Production-proven across 7 funds and 191+ counterparties.

How does CapitalBridge prevent the AI from hallucinating covenant numbers?+

Three layers. Tools, not generation: the MCP exposes specific read tools that return structured data from the same services the dashboards use, not free-text inference. Citation required: every figure carries a row identifier and a timestamp, so a credit officer can click through to verify it. Dashboard parity: the MCP and the web UI read from the same query path, so a number in chat is always identical to the number in the portal.

How is this different from Cardo AI, CovenantIQ, Allvue Nexius, or Octus CovenantAI?+

The incumbents (CovenantIQ, Allvue Document IQ, Octus CovenantAI) are developed-market focused, with a clean API to accounting systems but no MCP and no impact-covenant support. Cardo AI is AI-native portfolio analytics for private debt and securitisation. CapitalBridge competes on a different wedge: EM/DFI focus, impact covenants, facility-agreement variety, and MCP-native conversational delivery. Compare directly at vs Cardo AI and vs CovenantIQ.

How long does it take to connect Claude Desktop to my portfolio?+

Roughly 30 seconds. Open Claude Desktop, add the CapitalBridge MCP connector, sign in with Microsoft, and your real-user permissions flow through. No tokens to copy, no API keys to manage. The connector is read-only and audit-logged. Once connected, you can ask your portfolio anything Claude can reason about with the data: borrower watchlists, covenant reconciliation, deterioration analysis, impact-covenant reporting.

What questions can I ask the CapitalBridge MCP?+

Typical questions production users ask: "What's deteriorating across the book?" (a row-cited watchlist), "Covenant health for [borrower]" (per-facility status, headroom, last submission), "5 worst borrowers by compliance score" (ranked watchlist), "Chart the DSCR trend for [borrower]" (historical series), "Reconcile the configured covenant for [facility] against the signed agreement" (drift flag with citations to both the database and the document page), and "Which borrowers are missing this quarter's ESAP submission?" (overdue list). The tools are read-only and grounded in the same services that power the dashboard.