A reusable intelligence-and-action layer for your applications.
Cambiary sits between AI models and your business systems. Your app exposes context, tools, permissions, and workflows through two small seams; agents then plan, collaborate, act through approved tools, ask for approval before anything sensitive, render interactive results, and stay fully auditable — inside the app your users already use.
Embed the <agentic-panel> web component; pass a references-only "where am I" payload. The model never receives raw records.
validateToken · getContext
getPermissions · executeTool
pushNotification
Your app is the body. The platform is the brain.
The platform holds zero application-specific code — it never needs to know what your system is. You connect it through two generic seams, with no changes to the platform core.
HostContext — the web component
Drop the <agentic-panel> web component into any modern web stack — Angular, React, Vue, Razor/MVC, or plain HTML — and pass a references-only "where am I" payload. The model never receives raw records this way, and the host token is passed via config, never localStorage.
HostAdapter — five methods
Implement five methods — validateToken, getContext, getPermissions, executeTool, pushNotification — in-process in TypeScript or out-of-process over HTTP from any language. Your adapter is the authority; the platform issues no identities of its own.
Governance is an invariant, not a setting.
The AI runtime can never execute a side-effecting tool directly — it must call back through one governance gate, which runs permission, policy, approval, and audit on every action, even across the service boundary. There is no bypass.
Effective permission is the intersection of what the user, the agent, and org policy each allow — fail-closed, deny-by-default — and every decision is written to an append-only audit trail using references only, never raw record content.
Read-only, safe-create, controlled-write, destructive — the tier decides auto-run vs. human approval.
Agents propose a plan and wait; sensitive actions need an explicit yes; destructive actions are blocked.
Approvals are scoped so authority can't leak across agent handoffs — permissions are re-checked each time.
Authored instructions can't escalate privilege; injection attempts are evaluated at the boundary.
The parts, and how they fit.
Every capability is built and tested. See the full feature set →
Governance gate
One choke point for every side effect — permission ∩ policy ∩ approval ∩ audit, no bypass.
Context engineering
Minimization, boundary enforcement, freshness, and output provenance.
Multi-agent orchestration
A coordinator plus specialists, isolated state, and audited handoffs.
12 card types
Plans, approvals, results, comparisons, tables, and mermaid diagrams — no custom UI.
Voice surface
Talk to agents over OpenAI Realtime; approvals still require an explicit yes.
Admin & authoring
No-code agents and skills, teach mode, and drafts that stay drafts until you publish.
Analytics
A GA4 conversion funnel with a hard PII blocklist and opt-out.
Model routing
Task-aware routing across OpenAI models under allowlists; fail-closed. OpenAI today.
Plan → approve → act → record.
Plan
The agent proposes a plan card and waits — nothing sensitive runs yet.
Approve
A person approves, rejects, modifies, or supplies context; the gate re-checks permissions.
Act
Only through approved tools, and only within the intersection of user, agent, and org policy.
Record
An append-only audit entry captures the decision — references and IDs only, never raw content.
See it running in an app like yours.
Two seams, one governance gate, a live workflow — we'll walk your team through it using an example close to your stack.