An assistant that answers from your collection - and won’t make things up
Guide is museum intelligence wired to your actual records. It works across Collections, Media, and Content - searching your catalog before it answers, citing where the answer came from, and saying “that isn’t documented” rather than inventing an accession number. Included with every plan.
One brain, every app
Guide is trained on museum procedures, CDWA, NAGPRA, conservation literature, the NPS Museum Handbook, Getty vocabularies, and more - and it knows your records. The same Guide follows you across the platform, adapting to whichever app you are working in.
How Guide works
Guide is a multi-agent system, not a single chatbot. A lead agent reads your question, decides what kind of help you need, and delegates focused tasks to specialist agents - each with a narrow, server-enforced toolset for their domain. Specialists never delegate further, and every tool call is logged with its persona, arguments, and result.
Registrar
Cataloguing, vocabulary lookup, object lookup, record history.
Loans Registrar
Workflow status, overdue items, location and storage, contacts.
Conservator
Condition assessment, conservation literature, condition report drafts.
Rights Specialist
Rights interpretation, reproduction policies, contact lookup.
Curator
Exhibition planning, related objects, vocabulary research, web sources.
A planner persona decomposes multi-step goals into ordered plans before any specialist runs - so you can review the work before it starts.
Where Guide shows up
Same museum-trained model. Different context for each app on the platform.
Guide in Collections
Cataloguing help, vocabulary lookup, workflow step prompts, and record-aware search.
- “Which AAT term should I use for this object type?”
- “What fields are required to start an outgoing loan?”
- “Find every object on loan due back this quarter.”
Guide in Media
Asset discovery, rights interpretation, metadata help, and preservation event explanation.
- “Find images visually similar to this one.”
- “Which assets have rights expiring in the next 90 days?”
- “Explain why this file is flagged in the format registry.”
Guide in Content
Authoring help that pulls from your collection - page suggestions, related objects, and SEO guidance.
- “Suggest objects to feature for a textile exhibition page.”
- “Draft an exhibition guide block from these accession numbers.”
- “Rewrite this page intro for a younger audience.”
Grounded answers, not invented ones
Guide is built to be useful in a working museum. That means searching your actual records before answering, citing standards literature when relevant, and knowing when to defer.
Searches your records first
Before answering, Guide queries your collection, media, and content. No hallucinated objects or invented accession numbers.
Cites the source
When Guide draws on standards literature, it points to the specific procedure, regulation, or guideline so you can verify it.
Knows its limits
Trained to defer on legal determinations, conservation treatment decisions, valuations, and institutional policy. It guides - it does not decide.
Permission-aware
Guide respects the same field-level permissions as the rest of the platform. If a user cannot see a field, neither can Guide on their behalf.
Trained on the procedures you actually run
Guide ships with step-by-step playbooks for the procedures museums run every day. Ask "how do I X?" and Guide answers from the playbook - then takes you to the right page in one click.
Custom playbooks for institution-specific procedures are on the roadmap.
Drafts, not direct edits
Guide cannot write to your records. Every write tool produces a draft that flows through review and approval - never a silent change.
No silent writes
Guide can propose a condition report, draft a loan document, or stage a metadata update - but the change is a draft in your approvals inbox until a human accepts it.
Citations are validated, not generated
Every reference Guide cites in a draft is resolved against your records or the standards corpus at creation time. Hallucinated citations are rejected before the draft ever reaches your reviewer.
Sensitive records gate writes
Records flagged sensitive at the institution level block Guide’s write tools by default. Reviewers see why a draft was blocked and can route it to the right person.
Multi-step plans live in a durable inbox with Needs you / Active / History views, so a paused loan check or NAGPRA review never gets lost between sessions.
What Guide knows
Guide is tuned on the standards and records museum professionals rely on every day - and it grounds every answer in your own catalog.
25+
museum standards and authoritative sources
Museum procedures
Object entry, acquisition, loans, movement, cataloguing, deaccession, condition, and conservation.
CDWA
Categories for the Description of Works of Art - field-level guidance on what to record and why.
NAGPRA
Federal regulations, consultation requirements, notification procedures, and repatriation workflows.
Conservation science
CCI conservation bulletins, NPS Museum Handbook, condition terminology, and environmental standards.
Getty vocabularies
AAT, ULAN, and TGN with context on preferred usage and hierarchical relationships.
Rights & reproductions
Copyright assessment, reproduction policies, rights statements, and licensing guidance.
Digital preservation
OAIS framework, NDSA levels, format sustainability, and migration planning.
Insurance & valuation
Valuation methodology, federal indemnity programs, and claims procedures.
Extend Guide with your own knowledge
Madrona Guide speaks the Model Context Protocol (MCP) - so you can connect your own knowledge sources and tools without backend code changes or custom model training.
- Institutional style guides and curatorial policies
- Partner registries, consortium databases, and aggregator APIs
- Specialized vocabularies for region-, period-, or media-specific collections
Live infrastructure, expanding capability
Guide Studio
From asking to doing
Studio is the layer where Guide stops describing procedures and starts executing them - multi-step plans, draft generation, citation validation, role-based approvals, and a complete audit trail. Seven guided procedures are live in Madrona today, with more shipping one at a time.
Shipped today
- Plans Inbox with Needs you / Active / History views
- Draft-and-approve across every write tool - no silent writes to your records
- Seven procedures end-to-end: loan-in, accessioning, loan-out, deaccession, conservation treatment, media rights-to-publish, and reproduction requests
- Condition reports as drafts (Conservator specialist)
Next
- More guided procedures, added continuously
- Editable draft forms - adjust referenced values before approving
- Pre-approval trust tiers and write-surface safety checks
On the horizon
- Batch procedures - run one plan across many objects at once
- Branching procedures that fork on NAGPRA status, donor anonymity, and lender type
- Event-driven obligations and templates that learn your institution's preferences
Studio is a paid capability built on top of Guide. Pricing will be announced at launch.
The same Guide, on your public website
An embeddable widget puts Madrona Guide on your museum’s public site so visitors can ask about your collection, exhibitions, hours, and accessibility - in their own language, around the clock.
Locked-down visitor persona
Public Guide cannot see internal records or staff documents. Source grounding is required: if Guide does not have a documented answer, it says so.
Prompt-injection hardened
Visitor input is stripped of prompt-injection patterns before reaching the model, with strict per-session rate limits.
Embeddable widget
A small script tag drops Guide onto your site. Visitor questions become a source of insight into what audiences actually want to know.
Museum-trained reasoning, grounded in your records
Guide pairs a museum-specific model - tuned on tens of thousands of museum Q&A examples drawn from museum cataloguing standards, Getty vocabularies, NAGPRA, and conservation and rights practice - with retrieval over your live collection. Object numbers, creators, dates, and links are checked against what is actually in your records; ungrounded claims are blocked, not guessed. Your queries and records are never used to train anyone’s model.
Run Guide your way
Guide’s reasoning runs on the inference you choose - top-tier cloud, your own GPU, or a museum-specific model entirely on your own hardware. Same Guide, your call.
Top-tier cloud
Powered by Anthropic Claude for the highest-quality reasoning, with tenant isolation and a full audit trail on every interaction.
On your own hardware
Run a museum-specific model on a single on-site server - collection data and visitor conversations never leave the building, and Guide keeps working if the network drops.
Fixed cost at any scale
A public kiosk answering thousands of visitors a day costs the same as one answering ten. Local inference turns a usage-metered bill into a predictable fixed cost.
Serverless GPU
Burst to dedicated GPU for local-grade privacy without owning hardware.
Madrona Guide is also available as a standalone product. Upload your documents, embed a widget, no collection migration required.
See Guide in action
Schedule a demo to see how Guide works inside Collections, Media, and Content - and what Studio will add when it ships.
Request a demo