Extend
What you can do over MCP: every app feature and its tool
TL;DR
Over MCP, an agent connected to Agent Planners can create and change what a person does in the app, through the same rules the app's pages use. That is 17 kinds of workspace content: schedules, agent loops, projects, alert rules, dashboards, Forbidden Rules, the brand profile, skills, memory, assets, folders and share links. Changes a person must decide (a deleted file, a link anyone can open) wait for an approval, and a Forbidden Rule or brand-profile change always waits for an admin, whatever the key's mode.
Feature by feature
| Feature | Create | Read | Change | Delete | Notes |
|---|---|---|---|---|---|
| Tasks | create_task | get_task, list_tasks | send_task_message | stop_task | rev925: send_task_message is the app's message box (a follow-up, correction or answer — queued while a run is in flight; completed steps are not repeated) and stop_task the Stop button. get_task carries outcome (completed / partially_completed / blocked + what the person must do / failed) and what waits for approval. |
| Schedules | internal__create_schedule | internal__list_schedules | internal__update_schedule, internal__run_schedule | internal__delete_schedule | Run now is internal__run_schedule (a real run; each write in it waits for a person). |
| Agent loops | internal__create_loop | internal__list_loops | internal__update_loop | internal__update_loop | Created with auto-approve off; the key's creator must be an admin or editor. Stop (update_loop action stop) ends a loop. |
| Projects | projects__create, projects__generate_plan, projects__add_item | projects__list, projects__get, projects__status, projects__activity | projects__update_item, projects__activate, projects__pause, projects__resume, projects__refresh_numbers, projects__wake_overseer | projects__archive | Deleting a project for good is in the app (archive / restore over MCP). |
| Alert rules | internal__create_alert_rule | internal__list_alert_rules, internal__list_alert_history | internal__update_alert_rule | internal__delete_alert_rule | |
| Dashboards and reports | internal__create_dashboard, internal__emit_web_page, internal__export_report | internal__list_dashboards, internal__export_dashboard | internal__update_dashboard | internal__delete_dashboard | rev926: an external agent saves its own page with internal__emit_web_page saveAsDashboard (+ refreshQueries for automatic updates, restrictTo to keep one slice); update_dashboard publishes (a person approves), unpublishes, refreshes (regenerate), sets the refresh schedule and email recipients; export_dashboard downloads CSV / HTML / JSON. |
| Forbidden Rules | internal__create_forbidden_rule | internal__list_forbidden_rules, internal__draft_forbidden_rule, internal__test_forbidden_rule | internal__update_forbidden_rule, internal__set_forbidden_rule_enabled | internal__delete_forbidden_rule | Admin only. rev924: draft_forbidden_rule turns a sentence into a draft (nothing saved) and test_forbidden_rule runs the page’s “Test this rule” (nothing sent). A new rule is saved OFF; editing, turning on or off and deleting wait for an admin's approval on every key mode. |
| Brand & guardrail profile | In the app | internal__get_brand_profile | internal__update_brand_profile | In the app | Admin only; a change waits for an admin's approval on every key mode. |
| Skills | internal__save_skill, internal__sync_skills, internal__create_skill_from_url | internal__list_skills, internal__get_skill, export_skills | internal__update_skill, internal__sync_skills | internal__delete_skill | |
| Memory | internal__save_memory | internal__search_memory, internal__get_account_memory | internal__update_memory, internal__update_account_memory | internal__delete_memory | rev924: the agent also keeps what a person tells it (a follow-up's correction, a rejected write's reason) — a lasting rule as a long-term preference, anything else as a short-term note for that task only. The Memory page shows each row's source, age and term; a person edits or deletes it there or with update_memory / delete_memory. |
| Email workflows (mail rules) | internal__create_mail_rule | internal__list_mail_rules | internal__update_mail_rule | internal__delete_mail_rule | rev924: admin or editor; every change waits for approval like the agent's other standing automations. Reply-in-thread stays off unless asked. |
| Assets | internal__upload_asset_from_url, internal__upload_local_file, internal__start_local_upload, internal__finish_local_upload, internal__generate_image | internal__list_assets, internal__get_asset | internal__update_asset, internal__tag_assets | internal__delete_asset | Deleting a file waits for a person. |
| Asset folders (tags) | internal__tag_assets | internal__list_asset_folders | internal__update_asset_folder | internal__update_asset_folder | A folder is a tag: removing it keeps the files. |
| Share links | internal__share_folder, internal__share_asset | internal__list_folder_links | In the app | internal__revoke_folder_link | A link anyone can open waits for a person. |
| Composed tools | internal__define_composed_tool | internal__list_composed_tools | internal__define_composed_tool | internal__delete_composed_tool | |
| Web apps (worker apps) | internal__deploy_worker_app | internal__list_worker_apps, internal__get_worker_app | internal__promote_worker_app | internal__delete_worker_app | |
| Write approvals | In the app | list_write_requests, raw_request_status | In the app | In the app | Deciding an approval is a person's, in the app or from the email — a key never approves its own writes. |
| Audit log and usage | In the app | get_audit_log, get_usage, get_organization | In the app | In the app | |
| Connected accounts | google_ads__create_client_account, ga4__create_property, dv360__create_advertiser | list_accounts | internal__refresh_accounts | In the app | Connecting, re-authorizing and disconnecting are in the app (a sign-in at the platform). rev922: Refresh now is internal__refresh_accounts (it only adds and updates accounts); creating a Google Ads client account, a Google Analytics property or a DV360 advertiser waits for an admin's approval on every key mode, and the account list is refreshed after it. |
Tool names are the MCP names. When your client lists tools on demand (search_tools / call_tool), search for the feature by name — for example "forbidden rules" or "pause loop".
How the rules carry over
- Same checks as the page. Each tool calls the code the app's page calls: a Forbidden Rule over MCP is validated, saved OFF and audited exactly as on /forbidden-rules.
- Who the call acts for. An API key acts for the person who created it (an OAuth connection for the person who approved it), with that person's role now — admin-only features stay admin-only.
- Writes follow the key's mode. An ordinary change applies at once or waits for approval as the key is set. Changes a person must decide wait for an approval unless an admin set the key to apply them at once (or it is the workspace's own app). A guardrail change (a Forbidden Rule edited, turned on or off or deleted, the brand profile) always waits and is approved by an admin only.
- A key narrowed to some platforms sees and changes only the content about those platforms.
What stays in the app on purpose
| Feature | Why |
|---|---|
| API keys | A key could otherwise mint keys with more access than itself. |
| OAuth apps and connected apps | Registering an app or approving a connection is the consent step — a person's, signed in. |
| Billing, plans, credits and payment | Money moves only through a signed-in person's checkout. |
| Connecting, re-authorizing and disconnecting platforms | Each needs the person's own sign-in at the platform. |
| Members and roles | Who belongs to the workspace is managed in the organization settings. |
| Workspace settings (retention, notifications, spend limit, agent models) | They change limits that every key and every person works within. |
| Deciding approvals | A key never approves its own writes; a person decides in the app or from the email. |
| Suggested next steps (accept / edit / dismiss) | rev924: they appear under a finished task's report for the person reading it; over MCP the same step is simply asked for (e.g. internal__create_alert_rule). |
Frequently asked questions
- Can an MCP key turn off a Forbidden Rule?
- It can ask: the change is queued and an admin of the workspace approves or rejects it. A key never turns a rule off on its own, whatever its mode.
- Can an MCP key approve its own changes?
- No. Approvals are decided by a person in the app or from the approval email.