stable key, name, category
Let an AI agent prepare a change for review
Agent Bridge is the governed AI connection for Capell. It exposes the same capability layer used by Capell's AI Command Centre through MCP: an external AI client can discover intents, examples, and supported command shapes, validate structured inputs, find pages and sites by human-readable references, and request registered operations without receiving broad database or admin access.
Discoverable schemas Scoped authority Preview before mutation
Agent Bridge exposes a typed catalogue of Capell operations through MCP. The external client can discover capability names, categories, descriptions, intents, examples, input and output schemas, risks, scopes, package requirements, preview support, confirmation rules, and audit events before choosing a tool.
The tokenless public endpoint is limited to allowlisted AI Creator catalogues, interview, schema, and validation. The authenticated endpoint can list, run, and confirm site capabilities within token scope and current policy. A mutation receives a non-destructive preview, a separate matching confirmation, another policy check, and a secret-scrubbed audit record.
This external-agent path runs no inference in Capell. Optional AI Creator inside the admin is a separate provider-backed workflow. Both paths share structured site validation and keep generated sites private until a person confirms the exact application plan.
Before you adopt the AI packages
Both packages require Capell Membership. Use the live demo to inspect the public workflow first, then test the package path away from production. Keep a human responsible for provider terms, review, confirmation, and the final production release.
- Beta
- Maturity
- Agent Bridge and AI Creator are Beta in the marketplace.
- Membership
- Access
- Both are paid packages available through Capell Membership.
- Review first
- Production
- Pilot away from production and keep a person accountable for approval.
- Live demo
- Try it
- Inspect the public command flow before choosing a package.
Beta packages are suitable for evaluation and controlled pilots. Treat production use as a deliberate team decision.
Natural-language intent as a typed operation
Discover
List visible capabilities with names, descriptions, categories, intents, examples, schemas, risk, and confirmation rules.
Suggest
Match the user's words to likely capabilities without treating a suggestion as permission to execute.
Resolve
Find an accessible page, site, or package from its title, slug, URL, domain, name, or registered key.
Authorise
Check token scopes, current user policy, site access, installed package state, and native workflow readiness.
Preview
Return the exact non-mutating impact, warnings, target, and expected result for review.
Confirm
Bind approval to one user, capability, payload, preview, expiry, and audit event, then allow one execution.
Public building rules and protected authority
- Public read,
ai/capell/read: list the allowlisted theme, page-type, section-type, layout, interview, schema, and validation capabilities. It has no preview, export, build, apply, tokens, private records, page administration, or package activation. - Authenticated,
ai/capell: list site capabilities and run or confirm those allowed by the Agent Bridge token, current policy, package state, and site context. - A valid input is not authority. Every capability carries its own risk and confirmation contract, and policy is checked again immediately before execution.
Pre-invocation capability details
Capability discovery returns enough metadata for autocomplete and agent planning while keeping package internals private, including the direct-command shapes each capability accepts. Core page, site, publishing, package, and system operations live beside extension-contributed capabilities such as AI Creator. Installing or disabling an extension changes the catalogue and its command shapes without patching Agent Bridge.
description, intents, examples
input and output schemas
scope, server, policy ability
risk, preview, confirmation
required package and audit event
Website-lifecycle tools without blanket admin access
Start with read capabilities, then grant only the write scopes needed for the delivery workflow. A token can discover a broader catalogue than it may execute, and the capability still applies current user and site policy.
Questions
Pre-token authority questions
Does Capell choose or run the external model?
No. The MCP and Agent Bridge path runs no inference. The external client supplies the reasoning, credentials, model choice, usage cost, and retention terms. Optional in-product AI Creator is separate and uses the provider your team configures through the AIOrchestrator package.
Can an agent copy another website?
No. Page cloning is limited to content already inside the authorised Capell site. AI Creator may use abstract reference notes, but its prompt and validation rules reject scraping, close reproduction, impersonation, third-party branding, logos, and copied trade dress.
Can a voice request install code from the internet?
No. Package activation is limited to Composer-present, registered Capell packages. The preview shows declared impact before confirmation. Agent Bridge does not offer arbitrary shell or Composer execution.
Does a successful preview guarantee the operation will run?
No. The preview can reject an unready request. A successful mutation preview still needs the matching confirmation, and policy is checked again at execution. Tokens expire, are bound to the user and payload, and can be used once.
What reaches the public website?
Only approved render data. Prompts, token scopes, permissions, internal record identifiers, field paths, audit entries, signed editor URLs, model metadata, and private previews stay outside public output.
Loading section
Next step
Progressive agent authority
Start with discovery, then add each additional authority deliberately. Connect the public read endpoint first. Add a narrow token only after the team can explain its scopes, previews, confirmations, and audit review.
Use a real task as the acceptance test: find a named page, inspect it, preview a private clone, reject one unsafe request, and confirm one authorised operation. The in-admin AI-ready page shows how the same capability layer works for editors and voice input.
The agent-authority decision
Does the task have a named target and typed result?
- Proceed with Agent Bridge when
- The capability can resolve one permitted site, page, package, or operation and validate a bounded output schema.
- Keep the current workflow when
- The request depends on broad database access, arbitrary HTTP, a shell, or an open-ended instruction.
Does the task need write authority?
- Proceed with Agent Bridge when
- Read discovery is insufficient and the smallest required mutation scope can be named explicitly.
- Keep the current workflow when
- The public read endpoint or an existing Filament action already completes the job.
Can a person inspect the exact impact before execution?
- Proceed with Agent Bridge when
- The capability produces a non-destructive preview and binds one-use confirmation to the same user, payload, and operation.
- Keep the current workflow when
- The proposed effect cannot be made legible or policy must be bypassed to complete it.
Can a named operator own the audit and revoke access?
- Proceed with Agent Bridge when
- One owner can review secret-scrubbed events, rotate or revoke the token, and investigate rejected operations.
- Keep the current workflow when
- No person owns scopes, confirmation, incident review, or the external model and retention terms.