Manage {{product}} secrets, browser access, models, and administrator-authorized server folders. Use when the user asks to configure credentials or browser access, choose a coding provider, or attach an external folder.
Installs into .claude/skills of the current project.
Are you the author of Work Integrations?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/manishiitg-work-integrations)
---
name: work-integrations
description: Manage {{product}} secrets, browser access, models, and administrator-authorized server folders. Use when the user asks to configure credentials or browser access, choose a coding provider, or attach an external folder.
---
# {{product}} integrations
Inspect current state before changing it, and distinguish account-level setup
from selection for this project. MCP setup has its own `work-mcp` skill.
## Secrets
- Call `list_secrets` before creating, replacing, or deleting a secret.
- `set_workflow_secret` and `delete_workflow_secret` manage project-scoped
secrets in {{product}} (legacy tool names for the shared project store). There is
no account-level secret: a credential shared across projects must be a
shared Vault secret, copied by an administrator (see below).
- After `set_workflow_secret` succeeds, `$SECRET_<NAME>` is available to shell
tools immediately in the current chat and remains available in later turns.
Continue the requested work in the same chat; do not ask the user to start a
new chat or session. Verify availability without printing the secret value.
- {{product}} stores attached secret names in `workflow.json` under
`capabilities.selected_secrets`, using the AgentWorks workflow contract.
Secret values remain encrypted outside the manifest. Shared Vault credentials also require the executing user's current group permission. Respect the user's
selections in **Integrations > Secrets**; do not attach an unrelated credential.
- A read-only workflow or Crew reference never grants its secrets. To reuse a
credential across projects, an administrator can copy the source project/workflow secret with
`manage_global_secret(action="share", name="SOURCE_NAME", group_ids=["GROUP_ID"], vault_name="VAULT_NAME")`.
First use `manage_global_secret(action="list_groups")` to discover actual group IDs.
Use `list_secrets(source_workflow_path="EXACT_PATH")` when copying from another
accessible project, then pass the same `source_workflow_path` to the share action.
Explain that new shared secrets inherit Platform access by default, which administrators can revoke.
Ask which additional groups should receive access when the user's intent is unclear. Sharing keeps the source and existing attachments
intact, never overwrites a Vault name, and never returns plaintext. The two
copies rotate independently. The legacy `action="promote"` removes the source
copy; prefer share for requests to share a credential.
Alternatively create it through Vault > Secrets and grant use to an existing
group in Vault > Access > Secrets. Explicitly select that shared
name in each destination {{product}} project with
`update_project_global_secret_selection(action="select", name="NAME")`.
Call `list_secrets` first and use an exact name from `global.names`. {{product}} persists this allowlist in
`capabilities.selected_global_secret_names`; it never inherits newly created
globals automatically.
- Never print, echo, store in project files, or otherwise reveal a secret
value. Refer to secrets by name.
## Attached folders
- Use `list_work_folders` first. Attach only an existing path permitted by the
administrator, with a clear alias and the least access needed (`read_only`
unless writes are required).
- Use the exact stored ID when detaching. A newly attached folder becomes part
of the native CLI guard on the next turn; do not pretend the current turn's
process gained access retroactively.
## Browser and models
Browser mode and CDP settings live in the project's **Browser** view. Coding
provider and model settings live in **Setup > Models**: the provider is fixed
after the first message because it owns native conversation state, while the
model may still change. Use `agent-browser` for an actual browser task.
For provider or model configuration exposed through tools, discover the admitted
configuration tools and load their schemas. Never read or edit raw `config/`
files for LLM/provider configuration. The configured provider/model is the
runtime's identity; model self-identification does not override it.