A knowledge map you and your agents can align on. Captured and maintained by LLMs.
An agent reads the source's metadata and compiles it into plain, readable pages your team can read and correct.
Objects, fields, and types · Plus what they actually mean in your business. Not a column list; an explanation.
Validation rules, flows, triggers · The logic that decides what "qualified" really means. Surfaced, not buried.
Queries and joins that ran successfully, kept with the context they answered. The next agent starts from proof, not guesses.
# Example · One page from a Salesforce wiki
---
type: concept
source: salesforce
object: Opportunity
updated: 2026-06-04
---
# Opportunity
Deals in flight. "Qualified" is decided by the
Stage-Gate flow, not the StageName field · see
[automations/stage-gate-flow].
## Fields that matter
- Amount · committed value, USD. Set by flow
after quote approval; do not write directly.
- CloseDate_c · overrides CloseDate for
forecasting. Reports read the custom field.
## Joins that work
- Account 1:n Opportunity via AccountId
- Renewals: OriginalOpportunity__c self-join
## Rules
- Never sum Amount across stages < 3 (inflates
pipeline · validation rule VR-114)The wiki is maintained the way it was built · By an agent, diff-driven, on every change to the source.
Provision a connection and the first wiki compiles in hours · Objects, automations, the sales motion behind the schema.
Every agent reads the relevant pages before touching the source · Right fields, right joins, rules respected.
Successful queries and joins are harvested back in · So are your team's corrections.
One wiki per source today · One connected context as you add sources
A catalog is written by people and read by people · And it bit-rots between review cycles. The wiki is compiled and maintained by an agent, and its first reader is another agent. It captures what catalogs miss · The automations, the business rules, the joins that actually work · Because agents need those to answer correctly.
An agent does. Connect a source and it crawls the metadata · Schema, automations, flows, validation rules · And compiles readable pages. When the source changes, the wiki updates diff-by-diff. Your team reads, reviews, and corrects; nobody is assigned to "keep the docs up to date."
Yes. The wiki is plain, versioned, diffable pages · Not rows in a proprietary metadata store. Read it like documentation, correct it like a doc, review changes like code. It's readable, exportable, and yours.
Every agent working through the MarcoPolo workspace reads the relevant wiki pages before touching the source. It starts with the right objects, fields, joins, and rules in view · Instead of burning tokens rediscovering them · And different agents give answers that agree.
Any system you've set up in Connections. Salesforce is the deepest example today · Schema pages, an automations atlas, sales-motion synthesis · With the same approach applying across databases, warehouses, and SaaS systems.
Each connected source gets its wiki today, and the pages link to each other. As you add sources, that grows into one connected context across systems · The same customer traced through CRM, tickets, and billing. That cross-system graph is where this is headed.
Staleness is the design target. Updates are driven by diffs in the source · A new field, a changed flow · And health checks flag drift, broken links, and synthesis that no longer matches the schema. Maintenance is continuous agent work, not a quarterly project.
Inside your workspace, in your deployment · Your VPC or on-premise. Like everything in the workspace it's a versioned artifact · Portable files, no lock-in on your semantics.

Connect a source and get back a wiki of it · Objects, automations, the sales motion as it really runs. The fastest way to see what your agents have been missing.
