Captain – Multimodal RAG Search That Scales

TL;DR

TL;DR: Captain is a multimodal RAG platform that indexes video, images, PDFs, audio, and handwriting into a single searchable API — deployable in minutes with an MCP server for Claude Code and Cursor.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: runcaptain.com — verified product page content
• Source repository: github.com/runcaptain — verified via GitHub API (org: runcaptain)
• License: MIT — verified via GitHub API license.spdx_id
• MCP server: github.com/runcaptain/captain-mcp — verified via GitHub API
• HN launch thread: news.ycombinator.com/item?id=47366011 — “Launch HN: Captain (YC W26) – Automated RAG for Files”

What Is Captain?

Captain is a multimodal RAG (retrieval-augmented generation) search platform built by RunCaptain Inc, a Y Combinator W26 batch company. It connects to over 20 data sources — S3, SharePoint, Google Drive, Dropbox, Confluence, Slack, Gmail, and more — and automatically indexes their contents into a unified search API.

Unlike text-only RAG, Captain handles video, images, handwriting, audio, and scanned PDFs via auto OCR and vision-language model (VLM) description. A single POST query returns ranked results across all modalities with relevance scores.

The platform is API-first, SOC 2 certified, and ships with an MCP server (captain-mcp) that powers direct integrations with Claude Code and Cursor.

Key Features

From the runcaptain.com product page and GitHub repos:

Multimodal indexing:

• Auto OCR and VLM description for images and scanned documents
• Video frame extraction and description
• Audio transcription indexing
• Multi-lingual content support (verified against Tagalog test corpus)

20+ source integrations:

• Cloud storage: S3, Google Cloud Storage, Azure Blob, Dropbox
• Productivity: SharePoint, Google Drive, Confluence, Notion
• Communication: Slack, Gmail
• Custom: REST API upload, webhook ingestion

Search and retrieval:

• Hybrid search (keyword + semantic/vector)
• Re-ranking for precision
• Per-result relevance scores per modality
• Filter by custom metadata, file type, or date

Agentic tooling:

• MCP server for Claude Code and Cursor integration
• Python SDK (captain-py-sdk)
• TypeScript SDK (captain-ts-sdk)
• REST API (POST /v2/collections/{name}/query)

Benchmarks: Captain’s captain-mrag-bench repo evaluates retrieval quality on MRAG-Bench (ICLR 2025), a 16,130-image benchmark. Results show 81.3% ContentHit@5 — outperforming end-to-end GPT-4o + RAG (69.0%), Gemini Pro + RAG (65.9%), and Claude 3.5 Sonnet + RAG (63.6%) on the same benchmark’s retrieval subtask.

Setup Workflow

Step 1: Get an API Key

Sign up at runcaptain.com to receive your CAPTAIN_API_KEY and CAPTAIN_ORG_ID.

Step 2: Install the Python SDK

pip install captain-py-sdk

Or use the REST API directly.

Step 3: Index a Collection

import captain

client = captain.Client(api_key="your_api_key", org_id="your_org_id")

# Upload a file
client.collections["my-docs"].index_file(
    file_path="./whitepaper.pdf",
    metadata={"source": "marketing", "quarter": "Q2-2026"}
)

# Or ingest from a URL
client.collections["my-docs"].index_url(
    url="https://example.com/report.pdf",
    metadata={"source": "web"}
)

Step 4: Query Across Modalities

# Search across video, images, PDFs, audio, and text
results = client.collections["my-docs"].query(
    query="Find the white rose",
    top_k=5
)

for result in results:
    print(result.score, result.file_name, result.modality)
    # 0.95  video/preview.mp4  video
    # 0.75  images/handwriting.jpg  image
    # 0.30  docs/white-rose.pdf  document

Step 5: Connect via MCP (Claude Code / Cursor)

# Using the captain-mcp server
npx -y @runcaptain/captain-mcp

Configure your MCP client to point at the local server endpoint. Once connected, Claude Code and Cursor can query your Captain collections directly as a tool call.

Deeper Analysis

Where Captain excels:

• Heterogeneous document stacks (scanned PDFs + video +Confluence) where text-only RAG fails
• Scoped retrieval tasks where precision matters (internal knowledge bases, compliance archives)
• Rapid prototyping — no ops, no vector DB to manage

Where it may not be the right fit:

• Pure text corpora where a simpler embedding + FAISS stack is cheaper and faster
• Real-time streaming use cases (API is pull-based, not pub/sub)
• Very large-scale deployments that need full control over the embedding infrastructure

vs. established alternatives:

| | Captain | pgvector / FAISS | Azure AI Search | |---|---|---|---| | Multimodal (video/images) | ✅ Native | ❌ External VLM | ⚠️ Add-on | | Managed vector DB | ✅ | ❌ Self-managed | ✅ | | MCP server | ✅ | ❌ | ❌ | | SOC 2 | ✅ | Self-assess | ✅ | | Sources | 20+ | None (you build) | 10+ |

Practical Evaluation Checklist

• [ ] Sign up at runcaptain.com and get API credentials
• [ ] Run the quickstart with a sample PDF
• [ ] Test multimodal query (image + text in same query)
• [ ] Try the MCP server with Claude Code
• [ ] Benchmark against your current text-only RAG on the same corpus
• [ ] Check latency on your target corpus size (1K vs 100K files)

Security Notes

• SOC 2 Type II certified (independently audited)
• Role-based access control (RBAC) with custom metadata filtering
• Data never leaves your configured cloud storage unless you explicitly index it
• API key scoped to org ID — key rotation via dashboard

FAQ

Q: Does Captain store my files? A: Captain indexes metadata and VLM-generated descriptions. Raw files stay in your connected cloud storage (S3, SharePoint, etc.). You can also use client-side encryption before upload.

Q: What embedding model does Captain use? A: Captain uses a proprietary multimodal embedding model. The benchmark repo (captain-mrag-bench) evaluates against MRAG-Bench using their production API endpoint.

Q: How does it compare to a custom RAG pipeline with GPT-4o? A: On MRAG-Bench’s retrieval subtask, Captain scores 81.3% ContentHit@5 vs. 69.0% for a GPT-4o + CLIP-retrieved RAG baseline. The gap is larger on visual reasoning tasks (100% vs. 68.3% on high-precision queries).

Q: Is there a self-hosted option? A: Currently Captain is cloud-only (managed). Self-hosting is not documented as of July 2026.

Q: What is the pricing? A: See runcaptain.com/pricing. A free tier with limited queries is available. Volume pricing is per-query after the free allocation.

Conclusion

Captain fills the gap between text-only RAG and full multimodal AI pipelines. With 20+ source connectors, native MCP support, SOC 2 certification, and strong benchmark numbers on visual retrieval tasks, it is worth evaluating if your knowledge base spans video, images, or scanned documents — not just PDFs.

If you are currently using a text-only RAG setup and hitting precision walls on non-text content, Captain could replace several custom pipelines with a single API call.

• Try the free tier
• Read the docs
• View the MCP server