Knowledge
Department-organized policies, procedures, reference data, and glossaries. Authored once, agent-readable forever.
Knowledge layer
Agents are only as good as what they know.
Your procedures, policies, and rules, authored once in markdown and published where your credentialed tools can read them. The Knowledge base is built and usable in the portal today; the sections after it describe how a knowledge tree is organized, which is the shape a deployed agent will eventually read.
In the portal today
Your knowledge base, not a spec for one
This part is built. An organization creates departments, authors procedures and policies in markdown, publishes versions, and writes the rules its work has to satisfy — all inside the portal. Authoring requires a paid engagement and there is no checkout yet, so it is not self-serve today; reading is never gated.
Working
Authoring and publishing
- Departments you create in one call, each arriving with its overview page already written
- Documents authored in markdown with a live preview, filed under a fixed category vocabulary
- Draft, published, archived — and a version history you can read a previous revision from
- A per-department manifest generated from the rows on every read, so nothing can silently fall out of retrieval
Working
Rules and staleness
- Rule sets grouped by domain, each rule marked hard or soft
- Review dates on every document, surfaced as fresh, aging, or stale before anyone asks
- Agent visibility, so you can see which agents a department is exposed to
- Seeded rules ship with the structure and none of the numbers — your thresholds are yours to write
What reads it today: your own credentialed tools, over the REST API, the MCP server, and the Recurser CLI. Agents we host for you do not read it yet — that connection lands with hosted deployment. Rules you write are readable and reviewable; they do not yet block anything on their own. Flows is where a rule turns into a step that stops.
How it works
Markdown, but load-bearing.
Knowledge is authored as plain markdown with validated frontmatter, organized by department, and indexed for retrieval. That is deliberately boring: your compliance team can read and edit it, git can review it, and CI can reject it when it drifts out of spec.
The tree starts empty. We populate it with the commercial-lending content the agent needs, you author or override anything you want to own, and both live in the same graph.
There is no public install command and no module registry to pull from yet. Knowledge is provisioned with your agent deployment — talk to us if you want to author against the spec.
What makes it retrievable
- Registry
- Every document declares itself — id, department, owner, review date. Nothing is retrievable that is not registered.
- Taxonomy
- Department-organized folders with a fixed vocabulary, so two authors filing the same kind of procedure file it the same way.
- Retrieval map
- Routing from a question to the documents that can answer it, rather than embedding everything and hoping the vector search cooperates.
- Frontmatter CI
- Malformed or unowned frontmatter fails the build. A knowledge base that lies about its own metadata is worse than no knowledge base.
- Wikilinks
- Cross-document links form a graph the agent can walk — a covenant clause to the procedure that checks it to the policy that set it.
Taxonomy
Nine categories of agent knowledge
This is the filing system, not a storefront. Every category is a place where content lives in a deployed agent — what it knows, how it behaves, and what it is allowed to reach.
Knowledge & GovernanceWhat the system knows + how it stays in bounds.
Governance
KYC/AML rules, prohibited industries, retention policies, and audit-trail thresholds, written as machine-parseable YAML rather than prose an agent has to interpret.
Models
Per-task routing rules for the LLM router, so the model behind a task changes by configuration instead of code.
BehaviorWhat the agent does, says, and reaches for.
Skills
Named units of agent work with typed inputs, structured outputs, and explicit scope. A skill only becomes invocable once it lands in the capability registry.
Prompts
Persona and context prompts per department and role. Version-controlled, reviewed, and citation-ready.
Tools
Domain tool wrappers the workforce can dispatch — statement parsing, DSCR math, funder-appetite matching. Several are stubs today; the catalog below marks which.
IntegrationHow the agent connects to your stack.
Apertures
Connector definitions for the external systems a procedure needs to reach. Registered and permissioned per tenant; see /integrations for which connectors are live.
Hooks
Lifecycle interceptors — pre-decision validation, post-submit notification, audit emit. Composable, ordered, idempotent.
Plugs
Utility integrations — rate limiting, retry, cache, telemetry. Drop-in, with defaults that need no configuration.
We do not sell knowledge modules à la carte — there is no catalog and no per-module price. Knowledge is part of the agent deployment. What is separately counted is invocation of a registered capability, priced on /pricing.
Capability catalog
Skills are the verbs. Capabilities are the calls.
Knowledge describes what the agent knows. The capability registry describes what it can do — invocable, typed, priced. The same registry backs the REST endpoint and the MCP server.
2 functional · 5 registered stubs · 87 scaffolded placeholders
functional
Score CSG-1003 completeness
score_csg_1003_completenessfunctional
Generate credit memo
generate_credit_memostub
Analyze bank statements (per statement)
analyze_bank_statementsstub
Extract financials from documents
extract_financials_from_documentsstub
Verify business identity (inc. data pass-through)
verify_business_identitystub
Detect statement fraud
detect_statement_fraudstub
Calculate factoring borrowing base
calculate_factoring_borrowing_baseRead the badges literally. A functional capability does the work end to end. A stub has its request and response contract locked and its route wired, and returns a pending envelope rather than real analysis — stable enough to integrate against, not yet doing the work. Placeholders are scaffolding only.
Why this exists
Agents without knowledge are stochastic, not reliable.
Every credit decision, every funder match, every certified package needs a defensible trail back to a policy, a rule, a procedure. A structured knowledge layer makes that trail machine-parseable: the agent cites the document it sourced, and conflicts resolve against a documented hierarchy instead of whichever paragraph the model liked most.