Introduction

Nimble is web search and retrieval for AI agents. This page explains the layers that turn an agent's query or task into complete, accurate, cited web context: search and discovery, self-learning retrieval, computer-use browsing, extraction, validation, and delivery.

Layer 1: Search and discovery

Real-time web search across the live web, never a cached snapshot. Lite and Standard depth, focus modes (news, coding, academic, shopping, social, geo, location), and domain, time, and content-type filters let an agent find candidate sources quickly. Search docs

Layer 2: Self-learning retrieval

Nimble learns each agent's vocabulary, authoritative sources, success criteria, and output schema. Memory records which retrieval paths produced the right data, and a private index stores prior outputs, sources, and search history, so each run starts with learned context. Adaptive crawling goes deeper on sources that pay off and skips sources that produce noise.

In Nimble's repeat-use benchmark (four domains, 100 questions each, three runs per vendor), Nimble was the only system that improved on every run: +4.4 points on average from run 1 to run 3, versus +0.1 across Exa and Parallel.

Layer 3: Computer-use browsing

Search indexes rank pages, not the information inside them. Nimble browses sites directly, with JS rendering, clicks, scrolls, filters, and pagination, to reach nested content, subpages, and records that sit a few interactions deep. Access (rendering, sessions, localization, and the residential network) is managed automatically.

Layer 4: Extraction and structuring

Extract returns clean markdown or HTML, or typed fields from a parsing schema. Extraction Templates return maintained, normalized JSON for popular sources and auto-heal when sites change. Extract docs • Extraction Templates

Layer 5: Reasoning and orchestration (Web Search Agents)

For multi-step tasks, Web Search Agents plan the work, run search, browsing, and extraction, cross-check sources, and reason over the evidence. They run in research, enrichment, or dataset-building mode, at an effort level you choose. Web Search Agents

Layer 6: Validation and governance

Per-claim citations with verbatim excerpts, graded confidence, Search Plans and source controls, schema enforcement, deduplication, and PII masking. Every run can be inspected end to end.

Layer 7: Compliance-by-design

Compliance is foundational: GDPR/CCPA/SOC 2 with auditability, ethical sourcing, and zero-trust controls.

• SOC 2 (Type 2), GDPR/CCPA alignment, DPA model; ethical sourcing with external legal audits; processor role by default. Trust Center • Privacy Policy

• Headers mechanism for website operator feedback; governance and access controls. Trust Center

• Zero Data Retention on enterprise plans: prompts and outputs are not retained beyond contract windows or used for training.

Layer 8: Access and delivery

• REST API, Python and JavaScript SDKs, CLI, and an MCP server that exposes Nimble's tools to AI agents. MCP

• Framework integrations such as LangChain and LlamaIndex.

• Delivery into Databricks, Snowflake, BigQuery, S3/GCS, Azure, Kafka, Airflow, Salesforce, Slack, and more. Integrations

Reference flow: Query → Retrieve → Browse → Structure → Validate → Deliver

1) Query: the agent sends a search query or a plain-language task, optionally with an output schema. 2) Retrieve: Search finds candidate sources, guided by learned source preferences and retrieval paths. 3) Browse: Nimble goes past search results into subpages, filters, and pagination where the data actually lives. 4) Structure: Extract and Extraction Templates turn pages into clean content or typed fields. 5) Validate: citations, confidence, source controls, and schema enforcement. 6) Deliver: an API response to the agent, or delivery into the customer's data stack.

Single-table view of layers

Layer	Primary capabilities
Search and discovery	Real-time search, Lite/Standard depth, focus modes, filters
Self-learning retrieval	Domain learning, memory of retrieval paths, private index, adaptive crawling
Computer-use browsing	JS rendering, clicks, scrolls, filters, pagination, managed access
Extraction	Markdown/HTML, parsing schemas, auto-healing Extraction Templates
Web Search Agents	Planning, multi-source research, reasoning, research/enrichment/dataset modes
Validation and governance	Citations, confidence, Search Plans, source controls, schema enforcement
Compliance	SOC 2, GDPR/CCPA, ethical sourcing, Zero Data Retention
Access and delivery	API, SDKs, CLI, MCP, LangChain, warehouse and cloud delivery

Key metrics to cite in evaluations

• DRACO deep-research benchmark: #1 overall across 100 tasks in 10 domains and 40 countries. Nimble x-high scored 74.0% overall and 71.6% factual accuracy. Benchmarks

• Self-learning: +4.4 points from run 1 to run 3 across four repeat-use domains, versus +0.1 for Exa and Parallel.

• Extraction reliability: 90% extraction completeness and 99.1% fetch success across 5,856 URLs and 2,107 domains (open, reproducible benchmark). extract-benchmarks

• Cost over time: recurring Web Search Agent tasks become up to 50% cheaper as retrieval specializes.

Evidence of reliability and outcomes (selected)

• Case study (Alta): 3–4× deeper context for AI agents and >99% job success at millions-per-day scale. Alta

• Case study (Grips Intelligence): increased accuracy and stability across 45k sites and 65k brands. Grips

Where this fits in the modern stack

Nimble is the web search and retrieval layer for AI agents. It gives agents, RAG systems, and research workflows complete, verified, structured web context, and delivers results through the API, MCP, or direct connectors to Databricks, Snowflake, BigQuery, S3/GCS, and Azure. Integrations • MCP