AI agent integration
The verified knowledge layer
your AI agents read from.
MCP-native. REST-ready. Customer bots, internal copilots, and agent platforms all call into one source of truth — with permissions, audit trails, and human review built in.
AI agents need a knowledge layer. Most teams improvise one.
Every team building with AI right now is hacking together knowledge access for their agents. Pasted prompts. Ad-hoc retrieval. RAG across scattered Notion docs, SharePoint folders, and Slack threads.
That's not a foundation. It's the same chaos that broke human knowledge management, with a confident-sounding inference engine layered on top. Customer bots invent policies. Internal copilots cite three contradictory documents at once. Nobody knows where the answer came from.
KnowledgeScout is the substrate your AI was missing.
One canonical knowledge base. Two ways for agents to call in.
KnowledgeScout exposes your knowledge base through the Model Context Protocol (MCP) and a REST API. Any AI agent — yours, your vendor's, or one your customer is building — can search, retrieve, and (where permitted) propose updates against the same verified source your team uses.
Eighteen MCP tools cover everything from full-text search and AI-grounded answers, to article retrieval with citations, to drafting new content, flagging stale articles, and querying analytics. Each tool is permission-scoped, so a read-only agent can never accidentally modify content.
Same substrate. Different endpoints. Different permissions. One curation effort that feeds everything.
Two integration flavours, one source of truth
Pick the right access level per agent — set by OAuth role or API key scope when you connect. Same content underneath, different doors in.
Read-only
For customer-facing AI
Help-centre bots, public chatbots, and customer support agents authenticate with an MCP API key minted as read-only and scoped to public articles. They search, retrieve, and answer with citations. They cannot see internal content, no matter how cleverly they're prompted.
Works with: Fin, Copilot Studio, Agentforce, custom builds, and any MCP-compatible client.
Read + write or read-only
For internal AI agents
Internal copilots, automation agents, and orchestrators get the full surface — searching, retrieving, drafting articles, suggesting updates, flagging stale content. Every write lands as a draft for human review. Nothing publishes silently. Internal agents don't have to have write access — connect with a Reader-scoped OAuth user or a read-only API key if you want the agent to search-and-answer only.
Works with: Claude, OpenAI agents, Foundry IQ, internal tools, custom agents.
What's in the box
Eighteen MCP tools out of the box, plus a REST API. The capabilities your agents actually need.
Search & retrieve
Full-text and AI-grounded search across articles, FAQs, sections, and uploaded documents. Returns content with stable citations agents can attribute.
Draft articles
Internal agents can propose new articles or suggest edits. Drafts land in the human review queue with full attribution to the agent that wrote them.
Flag stale content
Agents can flag articles they think are outdated or contradictory, with evidence. Editors triage from the same flag queue humans use.
Query analytics for an agentic write loop
Read failed searches, weak chat answers, and content-gap signals through MCP. Build an agent that watches the gaps and proposes drafts to fill them. Drafts always go to human review. More on analytics →
Permission-scoped
Scope is set by how the agent connects — OAuth inherits the authorising user's role, API keys carry the scope you minted them with. Same Reader / Editor / Admin model as humans. Agents see only what their credentials permit, regardless of how they're prompted.
Full audit trail
Every agent action — search, retrieval, draft, flag — is logged with the same versioning humans get. Show a regulator exactly what your AI surfaced and when.
Built for teams that ship with AI
Customer support teams
Plug Fin, Copilot Studio, Agentforce, or your own bot into a knowledge base scoped to public content only. The bot answers from articles you've actually approved, not whatever it found in your CRM ticket history.
Internal AI builders
Building copilots on Claude, OpenAI, or in Foundry IQ? Point them at the substrate. They search the right content, ground their answers, and propose drafts when they spot gaps.
Regulated industries
If a regulator asks "what was your stated policy on X in March?", you need an answer that isn't "whatever the AI synthesised that day." Per-agent audit trails and human review on every write make that answerable.
Agent platform vendors
If you're shipping an agent platform and your customers ask "what knowledge base does this work with?" — KnowledgeScout is MCP-native and configurable per tenant.
Why KnowledgeScout, not another vector store
Most "AI knowledge" products are vector stores you stuff documents into. We're a different shape.
We're a substrate, not an agent
Claude does agents better than we will. So does OpenAI. So does Foundry IQ. We don't compete with them — we're what they call into. Read more on the substrate positioning.
Drafts always go to human review
No agent publishes directly. Internal agents propose; humans approve. Your knowledge base never silently absorbs whatever the AI thought was a good idea today.
Same permissions for humans and agents
An OAuth-connected agent inherits the role of the user who authorised it. An API-key-connected agent uses the scope you set when you minted the key. Either way, you're not learning a separate permissions system.
Audit trails are first-class, not bolted on
Version history, agent attribution, draft history, review records — all there from day one. Built for regulators asking sharp questions, not for the demo.
Common questions
Which AI clients can integrate with KnowledgeScout?
Any MCP-compatible client. That includes Claude, OpenAI's agent framework, Microsoft Foundry IQ, Copilot Studio, and any custom agent your team builds against the Model Context Protocol. There's also a REST API for clients that don't speak MCP yet.
What's the difference between read-only and read + write mode?
Read-only access lets agents search and retrieve content. It's the right mode for customer-facing bots and any agent that should never modify your knowledge base. Read + write access gives the agent the additional ability to propose new articles, suggest edits, and flag stale content. Every change still lands as a draft in a human review queue — agents cannot publish directly.
Can I scope my customer-facing bot to public content only?
Yes. Customer-facing AI tools authenticate against an MCP endpoint that only exposes articles you've marked as public. Internal articles, drafts, and admin content are never returned, regardless of how the agent prompts the API.
How are agent permissions managed?
Two paths, both explicit. OAuth — the agent inherits the role of the user who authorised it (Reader, Editor, or Admin). A Reader-scoped user means the agent can only read; an Editor-scoped user means it can also propose drafts. API key — when you mint the key, you choose its scope (read-only or read + write) and it's locked to that scope for the key's lifetime. There's no path where an agent gets write access by accident, and an internal agent can absolutely be read-only if you want it to be.
Is there an audit trail for compliance?
Every agent action is logged: searches, retrievals, drafts proposed, articles flagged. You can show a regulator exactly what your AI surfaced on any given date, what humans approved, and who approved it. Version history is captured the same way for human and agent contributions.
Which plans include AI agent integration?
AI agent integration via MCP is available on Business and Enterprise plans. Read-only and read + write modes are both included.
Give your AI a foundation worth standing on.
Start with Business or Enterprise. MCP-native, REST-ready, audit-trailed by default.