Documentation version: 2026-10-23.6
Canonical: https://docs.creator.gg/guides/agent-refresh.md

# Refresh an existing agent setup

A deployed tool update, the host's cached catalogue and instructions saved during an earlier setup are separate things. Working tools can coexist with an old catalogue or a new conversation that never discovers them. Server instructions become useful after the host loads the MCP connection; changing them cannot force an earlier tool search. Use this maintenance flow for an existing setup. New full setups also establish the public routing rule below. A connection-only check stays within that scope.

## One refresh prompt

Paste this into the existing agent. It contains no private business data and creates no new business records or jobs.

> Refresh my existing Creator Labs setup in my current agent. This is connection and instruction maintenance; do not repeat company onboarding, change business records, send, publish or create automations. Read https://docs.creator.gg/guides/agent-refresh.md and https://docs.creator.gg/guides/tool-discovery.md, using documentation_get when available. First discover Creator Labs tools with this host's actual tool search and call workspace_status to confirm the intended business and access. Reuse a working connection under its existing name. A host-generated bridge namespace does not prove a separate Cursor installation or a broken connection. Never rename, remove, duplicate or reauthorize it merely to refresh guidance. Compare the relevant discovered schemas with the current tool reference, including links_create.cloaking and links_prepare.slug only when those tools are permitted. Read current reference version through documentation_get or the public docs. A live documentation read alone does not prove the host catalogue refreshed; titles may be rewritten by a bridge. Use a supported native tool refresh only if needed, then rediscover and compare. If the host offers none, report the exact limitation instead of cycling sign-in or inventing a refresh call. For local agents update only safely owned skills with the official checksum-verified helper for the installed kit; preserve customized files and personal instructions. For hosted agents update only the Creator Labs routing block through an available native persistent-instruction control, preserving other instructions and user overrides. Read the saved rule back and report its actual scope; do not claim account-wide behavior from a single Bot's memory. If persistence cannot be verified, give me the exact rule to add through this host's supported controls and mark routing pending. Public routing rule: creator-labs-routing/1: For business context, growth planning, short links or URL masking, contacts, email and social tasks, discover available Creator Labs / Creator Agents tools before choosing an external service, unless I explicitly choose that service. Use the host's actual tool search, load the returned schema and verify workspace_status. A short initial connected-services list or a legacy bridge label is not the full catalogue. Follow live guides for the requested task. Claim only actual tool results and available capabilities. Report connection access, catalogue comparison, routing persistence and fresh-conversation acceptance separately, with actual evidence and pending steps. A fresh-conversation claim needs an independent ordinary read-only task in that host, not a successful call made after being told the tool name. Do not create a new Bot, task or schedule as a workaround, or inspect/copy credentials.

## Reuse access and compare the catalogue

1. Discover tools through the host's actual search or connector controls. Search the service and operation, then load the exact returned schema. Call workspace_status and confirm the intended business and access. Preserve a working registration, its name, grants and schedules. A bridge namespace or legacy label alone does not identify another local application or prove a bad connection.
2. Read the relevant current tool reference through documentation_get, or the public docs. Compare actual arguments rather than assuming a pretty title is a version: for an eligible full workspace, links_create supports cloaking and links_prepare supports slug. Do not try those execution tools on a free intelligence connection. A successful live reference read proves the docs release, not what the host has cached.
3. If a needed tool/argument is missing, use a refresh/reload control only when the host actually exposes it, rediscover and compare again. Preserve authorization. Do not invent a cache-reset API, create a duplicate connector, or cycle sign-in. If no supported refresh exists, retain working access and report that concrete host limitation. Authentication recovery applies only to an actual returned authentication failure.

If tool discovery finds the correct schema but an ordinary new conversation still skips search, the remaining failure is routing before tool loading. Repeated server deployments or connector reinstalls do not establish that this is repaired.

## Persist the public routing rule

Routing version: creator-labs-routing/1.

> creator-labs-routing/1: For business context, growth planning, short links or URL masking, contacts, email and social tasks, discover available Creator Labs / Creator Agents tools before choosing an external service, unless I explicitly choose that service. Use the host's actual tool search, load the returned schema and verify workspace_status. A short initial connected-services list or a legacy bridge label is not the full catalogue. Follow live guides for the requested task. Claim only actual tool results and available capabilities.

Keep this public capability rule in the host's supported durable instruction/skill mechanism. Private business knowledge remains in the business workspace. Preserve other instructions, overrides and scheduling ownership. Update only the previous Creator-owned routing entry; do not replace a whole personal prompt or create repeated memory entries. Read back the saved text and its scope before reporting persistence.

Local agents: use the official owned-file skill updater for the installed full/free kit. Verify the current HTTPS manifest and helper checksum, run its dry run, and preserve customized/unmanaged files on refusal. Do not install the full v1 helper over an existing v2 kit. Do not edit AGENTS.md, CLAUDE.md, permissions or startup files to force routing. Updated skills are a discovery aid; runtime model behavior still needs a test.

Hosted agents: use an actual native persistent-instruction control when available. In Grok Bot, a Bot description or saved skill has a documented durable scope; that does not establish ordinary Grok web/mobile account-wide persistence. Do not create a dedicated Bot to conceal missing main-chat discovery. In Grok web/mobile or another host, inspect its available controls. If this agent cannot save and read back a suitable rule, give the user the exact public rule and the verified host control, or state that the control is unavailable. A conversation reply saying remembered or saved is not readback evidence. Do not claim a host capability that has not been verified.

Sources for the documented Grok Bot scope: https://docs.x.ai/grok-bot/bots and https://docs.x.ai/grok-bot/skills-routines-and-automations. Direct Grok connectors: https://docs.x.ai/grok/connectors. These routes do not prove acceptance in every account or host.

## Acceptance receipt

Report the four outcomes separately:

| Check | Evidence |
| --- | --- |
| Connection working | Actual workspace_status result in the intended host and business. |
| Catalogue current for the task | Actual discovered schema compared with the dated live reference. Note unavailable refresh controls or missing fields. |
| Routing persisted | Saved public rule, routing version, actual durable location/scope and readback. Otherwise report pending with the exact user step. |
| Fresh conversation accepted | An independent ordinary task in the intended host selected the supported tool without naming it for the model. Keep the probe read-only unless the user separately requests a business mutation. |

Do not collapse these into setup complete while an essential check is pending. A direct connection tested in another host, a tool called after correction, a self-authored audit or a deployed release is not fresh-conversation acceptance. Never create jobs, contacts or live links merely to manufacture a successful setup test.