> ## Documentation Index > Fetch the complete documentation index at: https://docs.venture.caplia.ai/llms.txt > Use this file to discover all available pages before exploring further. # MCP Server > AI agents — Claude Desktop, Cursor, ChatGPT, Windsurf — that score pitch decks, manage deal flow, and read venture-capital pipelines through Caplia. **Caplia is an MCP server for venture-capital pitch-deck scoring and deal-flow management.** AI agents — Claude Desktop, Cursor, ChatGPT, Windsurf, or any custom MCP client — submit pitch decks, get back CRI quality scores and thesis-fit scores, list companies in the caller's pipeline, pull data-room contents, and poll async jobs as they complete. Built for VC funds, accelerators, angel networks, and corporate-development teams who want to drive their Caplia pipeline from inside the AI tools they already use, without copy-pasting between tabs. ## What you can do Concrete scenarios where Caplia MCP pays for itself: Drop a PDF into Claude Desktop, ask the agent to submit it to Caplia. The deck flows through Caplia's CRI scoring + thesis matching; the agent polls until results land and summarises them. "What's at the top of my pipeline this week?" — the agent calls `caplia_list_companies` and `caplia_get_company_scores` and ranks deals by CRI + thesis fit. "How well does this company fit our Climate Series A thesis?" — the agent fetches scores and walks through each scoring domain. The agent pulls company profile, traction metrics, data-room documents, and thesis scores into a structured memo, with citations back to Caplia. "Open the Tesla data room and find any term sheets." — the agent calls `caplia_list_company_documents`, identifies the term sheet, and fetches a signed download URL. Forward decks from your inbox to a small script that calls `caplia_submit_deck` per deck. Decks land in your pipeline pre-tagged with the company name your CRM already knows. ## FAQ Yes. The `caplia_submit_deck` tool accepts a PDF (base64-encoded), runs it through Caplia's intake pipeline — text extraction → company profile → **CRI quality score** + **thesis-fit scores** — and returns a `job_id`. Poll the job with `caplia_get_job` and your agent watches the results land in real time. Endpoint: `https://mcp.venture.caplia.ai/` (production) or `https://mcp-sandbox.venture.caplia.ai/` (sandbox). CRI (Caplia Readiness Index) is Caplia's proprietary 0-100 quality score for a company. It rolls up signals across team, market, product, traction, financials, defensibility, and execution, with per-domain breakdowns. The MCP server returns the full structure — overall score, per-domain scores, and the individual signals that drove each domain — so agents can explain *why* a deck scored the way it did, not just *that* it scored. Yes. ChatGPT supports remote MCP servers via custom GPT actions and through the ChatGPT desktop MCP integration. Point it at `https://mcp.venture.caplia.ai/` with an `Authorization: Bearer cap_inv_live_*` header. See [Configure your MCP client](#configure-your-mcp-client) for the analogous config — same JSON shape, just different settings UI. All three work the same way — they're MCP clients that accept a remote MCP server URL plus an Authorization header. Setup snippets for each are below in [Configure your MCP client](#configure-your-mcp-client). Both. Caplia is used by venture funds, accelerator programmes, angel networks, and corporate-development teams. The MCP server reflects whatever pipeline structure your Caplia account has — your views, your theses, your stages. There's nothing VC-specific in the protocol itself. Caplia's MCP server is free for any Caplia customer — there's no separate "API tier" or per-call charge in v1. Rate limits are enforced organically (a key that abuses the API gets revoked); we'll add explicit usage tiers if and when there's a real need. Two big differences. **First, scoring is consistent and explainable.** CRI uses the same scoring rubric across every deck, with traceable per-domain signals — not the LLM's mood that day. **Second, it lives in your pipeline.** Scored companies show up in your existing Caplia views, theses, and team dashboards. They're not one-off chat outputs that disappear after the conversation ends. Yes. Every MCP request runs under the calling user's row-level security in Postgres — same isolation as the web app. Cross-tenant access is structurally impossible. There's also a **sandbox environment** (`mcp-sandbox.venture.caplia.ai`, `cap_inv_test_*` keys) with isolated test data, so integrators can develop without touching production. Yes, use the [REST API](/api-reference/introduction) directly: `POST /v1/decks` for submission, `GET /v1/jobs/{id}` for polling. The MCP server wraps these same endpoints. Same keys, same scopes, same data. Mint them yourself: **Caplia portal → Settings → API Keys → Create new key**. Pick `read` for read-only integrations or `write` if you need to submit decks. Keys are shown once at creation; copy to a secret manager. See [Authentication](/authentication) for the full key format and security guidance. ## Endpoints ``` Production: https://mcp.venture.caplia.ai/ Sandbox: https://mcp-sandbox.venture.caplia.ai/ ``` Transport: **Streamable HTTP** (single POST endpoint, JSON-RPC 2.0). Protocol version `2024-11-05`. ## Authentication Bearer-token, same model as the REST API: ``` Authorization: Bearer cap_inv_live_<32 chars> # production Authorization: Bearer cap_inv_test_<32 chars> # sandbox ``` Mint a key in **Settings → API Keys** inside Caplia. See [Authentication](/authentication) for full details on key format, scopes, and security best practices. ## Tools 13 tools that map 1:1 to REST endpoints. All tools inherit row-level security — you only see data your account already has access to in the web app. ### Read tools (require `read` scope) | Tool | What an agent asks for it | | ------------------------------- | ----------------------------------------------------------------------------- | | `caplia_search` | "Find any companies in my pipeline named Tesla" | | `caplia_list_companies` | "Show me companies at the seed stage" | | `caplia_get_company` | "Tell me everything about company X" | | `caplia_get_company_scores` | "What's the CRI score and thesis fit for company X?" | | `caplia_get_company_metrics` | "Show me traction metrics for company X" | | `caplia_list_company_documents` | "What's in company X's data room?" | | `caplia_get_document_url` | "Download the pitch deck for company X" | | `caplia_list_theses` | "What investment theses does my team have?" | | `caplia_list_properties` | "What custom properties (e.g. award categories) can I tag a submission with?" | | `caplia_list_views` | "What pipeline views are configured?" | | `caplia_get_view_companies` | "Show me everything in the Monday screening view" | | `caplia_get_job` | "What's the status of the deck I just submitted?" | ### Write tools (require `write` scope) | Tool | What an agent asks for it | | -------------------- | ---------------------------------------------------------------------------------------- | | `caplia_submit_deck` | "Submit this PDF as a new deal, tagged Climate Tech, with the founder's contact details" | Max 50 MB per deck (base64-encoded). Other write endpoints (create company shell, attach documents, add notes) exist on the REST API and will be wrapped as MCP tools as customer demand surfaces. ## Configure your MCP client **Using Claude in the browser or on mobile?** Caplia isn't in Anthropic's connector directory yet, and the **Settings → Connectors → Add custom connector** flow on claude.ai currently only completes for OAuth-based servers. Caplia authenticates with a static API key, so adding `https://mcp.venture.caplia.ai/` there fails — typically with *"Couldn't register with sign-in service"*. Connect through [Claude Desktop](#claude-desktop) or [Claude Code](#claude-code) below instead; both send your API key with every request. ### Claude Desktop Edit `~/Library/Application Support/Claude/claude_desktop_config.json` (macOS) or `%APPDATA%\Claude\claude_desktop_config.json` (Windows). Claude Desktop reaches remote servers through the `mcp-remote` helper, which needs [Node.js](https://nodejs.org): ```json theme={null} { "mcpServers": { "caplia": { "command": "npx", "args": [ "-y", "mcp-remote", "https://mcp.venture.caplia.ai/", "--header", "Authorization: Bearer cap_inv_live_YOUR_KEY_HERE" ] } } } ``` Quit Claude Desktop completely and reopen it (not just the window). The Caplia tools appear under the plug icon at the bottom-left of any chat. ### Cursor In Cursor settings → **Features → MCP → Add new MCP server**: * **Name**: `caplia` * **Type**: `sse` (streamable HTTP) * **URL**: `https://mcp.venture.caplia.ai/` * **Header**: `Authorization: Bearer cap_inv_live_YOUR_KEY_HERE` ### Claude Code One command, no file editing: ```bash theme={null} claude mcp add --transport http caplia https://mcp.venture.caplia.ai/ \ --header "Authorization: Bearer cap_inv_live_YOUR_KEY_HERE" ``` Run `/mcp` inside Claude Code (or `claude mcp list`) to confirm. Add `--scope user` to enable it in every project. ### Codex Codex connects to remote MCP servers natively, reading the key from a named environment variable: ```bash theme={null} export CAPLIA_API_KEY="cap_inv_live_YOUR_KEY_HERE" codex mcp add caplia --url https://mcp.venture.caplia.ai/ --bearer-token-env-var CAPLIA_API_KEY ``` Run `/mcp` in the Codex TUI (or `codex mcp list`) to confirm. ### Windsurf, Cody, ChatGPT, generic clients Most MCP clients accept the same three things: 1. URL: `https://mcp.venture.caplia.ai/` 2. `Authorization` header: `Bearer cap_inv_(live|test)_` 3. (Optional) Server name: `caplia` Refer to your client's MCP documentation for the exact config-file location. ## Quick test from the terminal If you want to verify the server works before wiring it into an agent: ```bash theme={null} # Liveness + protocol check (no auth) — a real MCP server answers `initialize` curl -X POST https://mcp.venture.caplia.ai/ \ -H "Content-Type: application/json" \ -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"curl","version":"1"}}}' # → {"jsonrpc":"2.0","id":1,"result":{"protocolVersion":"2024-11-05","capabilities":{"tools":{}},"serverInfo":{"name":"caplia-api","version":"1.0.0"}}} # List tools (no auth) curl -X POST https://mcp.venture.caplia.ai/ \ -H "Content-Type: application/json" \ -d '{"jsonrpc":"2.0","id":2,"method":"tools/list"}' # Submit a pitch deck (requires write scope) B64=$(base64 -i deck.pdf | tr -d '\n') curl -X POST https://mcp.venture.caplia.ai/ \ -H "Content-Type: application/json" \ -H "Authorization: Bearer cap_inv_live_..." \ -d "$(jq -n --arg b64 "$B64" '{ jsonrpc:"2.0", id:1, method:"tools/call", params:{ name:"caplia_submit_deck", arguments:{file_b64:$b64, filename:"deck.pdf"} } }')" ``` ## Troubleshooting | Symptom | Likely cause | Fix | | -------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------- | | *"Couldn't register with sign-in service"* when adding a custom connector on claude.ai | claude.ai custom connectors require OAuth; Caplia uses API keys | Use the [Claude Desktop](#claude-desktop) or [Claude Code](#claude-code) setup above | | Tools are listed, but every call returns `Missing Authorization header` | Your client connected without sending the key — `initialize` and `tools/list` don't require auth, so setup can look successful even with a missing or mistyped header | Re-check the `Authorization` header (or `--header` line) in your config, then restart the client | | `401` / invalid key, even though the key looks right | Key and host environments don't match | `cap_inv_test_*` keys only work on `mcp-sandbox.venture.caplia.ai`; `cap_inv_live_*` keys only on `mcp.venture.caplia.ai` | | No Caplia tools in Claude Desktop after editing the config | App not fully quit, a JSON syntax error, or Node.js missing (`mcp-remote` needs it) | Validate the JSON, install [Node.js](https://nodejs.org), then quit Claude Desktop completely and reopen it | | Tool calls fail with a permissions error on writes | Key has `read` scope only | Mint a `write`-scoped key in **Settings → API Keys** | ## Limitations * **One write tool today** (`caplia_submit_deck`). Other REST write endpoints (create company shell, attach documents, add notes) will be wrapped as MCP tools as customer demand surfaces. * **No streaming responses.** Each tool call is a single request-response. Long-running operations (deck processing) are polled via `caplia_get_job`. * **50 MB max deck size**, base64-encoded. * **JSON output only.** Each tool returns raw JSON from the REST endpoint, serialised as a text content block. Agents parse it normally. ## Source Implementation: [`supabase/functions/mcp-v1/`](https://github.com/Caplia-Tech/caplia-api/tree/main/supabase/functions/mcp-v1). Tool definitions are inline in `index.ts`. Open a PR there to propose new tools or refine existing ones.