Grounded retrieval & evidence for AI.
Turn AI answers into grounded, source-backed outputs. NOVA DeepSource retrieves real evidence, structures it, and returns results tied to real sources instead of unsupported model guesses.
Available as MCP Server and Developer API — Starting at $149/mo
MCP clients:
Ground your AI in evidence, not guesses
Retrieves real evidence relevant to the query — not just the closest vector match. Every result is tied to a real source.
Every AI output comes with a trail back to the sources that informed it. Audit, verify, and explain any answer.
Results come back as structured data your product can build on — not raw text blobs. Evidence graphs, execution traces, and telemetry included.
Chain retrieval stages into pipelines with versioned steps. Each stage has typed inputs, outputs, and clear contracts.
Use DeepSource from ChatGPT, Claude, Cursor, and agent clients via MCP — or embed the API directly in your own product.
Both MCP and API access included in every plan. No separate subscriptions, no surprise charges.
Each is powerful alone. Together they multiply.
Compresses all 50 tools in your registry — but you're still sending irrelevant ones alongside the relevant ones. 85% savings, but 100% of the noise.
Retrieves the 5 relevant tools from your 50-tool registry — but they're still fat with tokens when they hit the LLM. Right tools, full price.
DeepSource finds the 5 that matter. Optimizer consolidates them into 1 super-tool. 10,000 tokens become 300. 97%+ savings AND better accuracy.
Save 20% when you bundle both products
Start small. Scale when you're ready.
For teams evaluating grounded retrieval
For production workloads
For teams with custom requirements
Save 20% when you bundle with NOVA Optimizer — see platform pricing
Instead of letting an LLM generate answers from its training data alone, grounded retrieval finds real evidence first, then gives the LLM that evidence to work with. The result is AI output tied to verifiable sources — not unsupported guesses.
DeepSource exposes its capabilities as MCP tools. Any MCP-compatible client — ChatGPT, Claude, Cursor, or your own agent — can call DeepSource to retrieve grounded evidence as part of its workflow. You also get a Developer API for direct integration into your own product.
Yes, and they're designed to work together. DeepSource retrieves only the relevant evidence, then Optimizer compresses it further. The combination delivers 97%+ token savings with better accuracy than either product alone. Bundle pricing saves you 20%.
DeepSource is accepting qualified teams now. Early access means you get direct access to the product and our team, help shape the roadmap, and get priority when new capabilities ship.
All outputs are structured JSON your product can build on directly. Results include evidence graphs, execution traces with per-stage telemetry, and source-backed data. No parsing unstructured text — everything comes back typed and ready to use.
DeepSource is accepting qualified teams now. Tell us what you're building and we'll get you set up.