Enterprise AI needs live, governed web data. This page details the control surface for Nimble's retrieval stack: how Web Search Agents, the Web Tools APIs, and the hosted MCP server are configured, constrained, and audited, so teams can connect AI systems to the public web with confidence.

What runs where

Web Search Agents are the managed research layer. Send a task in plain language, optionally with a JSON schema; the agent plans the work, searches and extracts across the live web, cross-checks conflicting sources, ranks them by authority, and returns a cited answer or a structured dataset (overview).
Web Tools are the composable data layer when your agent owns the reasoning: Search, Extract, Map, Crawl, and pre-built Extract Templates, all served through one SDK (documentation).
The Nimble MCP server (https://mcp.nimbleway.com/mcp) exposes the platform as Model Context Protocol tools over Streamable HTTP. It authenticates with OAuth 2.1 and PKCE by default, so users sign in through the browser and no key is pasted or stored; clients without OAuth pass an API key as a bearer token. Tools cover Search, Extract, Map, Crawl, Extract Templates, and Web Search Agents, every tool name prefixed nimble_ (MCP server).

The agent configuration is the policy surface

Each Web Search Agent is governed by explicit, per-agent configuration (agent configuration):

Field	Governance role
`skill`	A briefing that fixes what dimensions matter and what jargon means, so behavior is specified rather than emergent
`goals`	Checkable objectives the agent works through on every run
`sources`	Allow-lists, block-lists, and source hints, for example preferring official pricing pages over third-party summaries
`effort`	A dial from `low` to `x-high` bounding how much work each run invests
`output_schema`	A JSON Schema; answers arrive as structured data with every value cited
`use_case`	`research`, `enrichment`, or `dataset_building`

Pre-built templates ship with goals, source guidance, a skill briefing, and output schemas already tuned, including competitive-intelligence, due-diligence, lead-enrichment, and business-discovery.

Effort and cost governance

Every run carries an effort level that controls fan-out, research rounds, and source-verification depth, with one flat price per level: low at $0.025, medium at $0.10, high at $0.50 (the default), and x-high at $2.00 per run. Billing is runs times effort price, with compute, retrieval, and storage included; set a default on the agent and override it per run (Efforts). For Extract workloads, driver selection is automatic for performance with a manual override for cost control (pricing).

Search Plans and retrieval transparency

Auditable Search Plans show exactly what was searched, where, and why (pricing), and control how data is retrieved from specific sources. Web retrieval stops being a black box: engineers can inspect why information was retrieved and where each datapoint entered the system.

The trust report: per-claim auditability

Every Web Search Agent run returns a trust report alongside the output (Trust documentation):

Every claim is cited. Each statement in a prose answer, and each value in a structured one, links to the exact URLs it came from, with verbatim excerpts from the page.
Every claim is graded high, medium, or low under fixed rules applied the same way to every run: a primary source grades high, corroboration across two or more independent domains grades high, a single secondary source grades medium, and no usable citation grades low. The reasoning field states which rule applied.
Every source is classified as official, news, social, or academic, and ranked primary or secondary for the claim it supports.
Structured outputs carry evidence per cell. With an output schema, claims are keyed by JSON path. A null backed by citations is a verified absence and grades high; a value that fails schema validation grades low, so structural errors are always flagged. A dataset's overall confidence is the lower of its fill rate and its claim-trust ratio.
Trust is machine-consumable. Route low-confidence values to human review, accept high ones automatically, and render citations in your own product.

Run lifecycle and audit trail

Runs execute asynchronously with explicit states (queued, completed, failed, cancelled); poll for results or stream progress as server-sent events (API reference).
Agents are scoped to the workspace resolved from the API token.
Agent changes apply as JSON Patch operations, so every modification is an explicit, reviewable operation.
Deactivating an agent is a soft delete: it can no longer start new runs, but existing runs and their results remain retrievable.

Governing MCP access

OAuth scopes: nimble:read, nimble:write, offline_access, and openid. The consent screen grants permission to verify identity, search the web, extract data, and manage Web Search Agents. Tokens refresh automatically and can be revoked by disconnecting the client or from the Nimble account.
API keys: issue a separate key per project from Settings → API Keys, so one can be revoked without breaking the rest. In Snowflake the key lives in a Snowflake Secret; in Databricks, in a secret scope or a Unity Catalog connection.
Databricks: the Marketplace connection routes every Nimble call through Unity AI Gateway. Calls are recorded in system.access.audit with the calling user, connection, method, and status code, and USE CONNECTION grants control who can call Nimble. Payload logging, service policies such as a domain allowlist on nimble_search, and per-user rate limits are available in Beta.
Snowflake Cortex Code: admins can enforce the Nimble MCP server for an entire account through managed settings, and pre-approve read-only tools while requiring confirmation for long-running ones such as crawl.
Microsoft: enterprises can expose Nimble MCP from their own Azure API Center registry, optionally proxied through Azure API Management.

Data protection

Zero Data Retention (ZDR) is supported under enterprise agreements arranged with Nimble's sales team: prompts, input data, and outputs are retained only within the window defined in the contract, are never used for training, and are never shared with any other customer. Retention is set per data type, including whether the agent keeps memory and storage across runs (documentation).
PII masking is available by arrangement with Nimble's sales team.
Nimble is SOC 2 Type II audited and enforces a Zero-Trust Architecture with encryption in transit and at rest, RBAC, and audit trails (Trust Center). The full security posture is on Nimble Security, Compliance, and Trust.

Ethical sourcing

Nimble's IP addresses are ethically sourced and undergo an external legal confirmation audit with detailed reporting; Nimble participates in the Ethical Web Data Collection Initiative, and residential IPs are not used for misrepresentation (Trust Center). Services target publicly accessible web data, and customers must comply with applicable laws and Nimble's Acceptable Use Policy.

Nimble Web Search Agents and MCP: Governance, Controls, and Auditability