Built-in tools
Every operation the OS answers for a task itself, with its inputs, effects, gates, retry safety and errors.
Generated from the built-in tools registry. Do not edit by hand.
A task calls two kinds of tool. Vendor tools come from the MCP servers an enterprise binds on the Tools page: GitHub, Slack, Sentry and the rest. Built-in tools are answered by the OS itself: os.*, workspace.*, memory.* and web.*, the CEO chat’s own tools, and the verbs of a Computer-Use binding.
This reference lists every built-in tool. It is generated from the same definitions the model is served, so what you read here is what a run sees.
Declaring a tool
A task can call only the tools it declares, by exact name, in its tools:
{
"tools": [
"workspace.read_file",
"workspace.edit",
"workspace.git_commit",
"os.set_execution_name"
]
}
- A name in a built-in namespace (
os.,workspace.,memory.,web.) that is not listed here is refused when the task is saved (tool_unknown). So is a tool only the CEO chat offers (tool_not_run_callable): a run never answers one. - A task can pin a field to particular values:
<tool>#<field>=<value>,<value>. The model is offered only those values, and a call carrying any other is refused. - The model calls a tool with its dots replaced:
workspace.read_fileisworkspace__read_file. - Agents can read this reference as data:
GET /v1/tools/built-inandGET /v1/tools/built-in/:name, with a session or an API token scopedtools:read.
The tools
Read and write
Read tools change nothing. Write tools change something: a file, a label, a decision, a note in memory. A run credential can write only if its task declares an os.* tool that writes.
Write tools: workspace.write_file, workspace.delete_file, workspace.edit, workspace.run_command, workspace.git_commit, workspace.git_push, workspace.share_file, memory.train, os.set_execution_name, os.add_execution_link, os.wait_execution, os.stop_execution, os.set_gate_summary, os.decide_gate, os.ask_ceo, os.record_issue_outcome, os.decide_recommendation, os.mail_reply, os.mail_send, os.mail_contact_save, os.mail_set_label, os.mail_set_handled, os.mail_set_many, os.mail_mark_chased, os.mail_snooze, os.my_mail_set_label, os.my_mail_set_handled, os.my_mail_set_many, os.my_mail_reply, os.my_mail_send, os.my_mail_forward, os.my_mail_delete, os.my_mail_start_action, os.my_mail_schedule_action, os.my_mail_unschedule_action, os.my_mail_contact_save, os.my_mail_contact_delete, os.my_mail_signature_set, os.my_mail_address_signature_set, os.escalate_mail, os.mail_reply_draft, os.record_kpi_reading, os.set_enterprise_kpis, os.set_team_kpis, os.set_context, os.delete_context, kpi.record_reading, role.define_kpi, role.edit_kpi, os.retry_run, os.route_run, os.clear_run, os.escalate_run, os.write_plan, os.chat_write, os.chat_open, os.start_task, os.hand_off_task_brief, os.finish_waiting, os.restart_waiting, os.clear_blocker, os.create_role, os.update_role, os.delete_role, os.add_role_task, os.remove_role_task, os.add_role_member, os.remove_role_member, os.create_task, os.publish_task_version, os.set_task_run_status, os.set_task_schedule, os.remove_task_schedule, os.set_task_input_values, os.write_context, memory.upload, os.create_team, os.update_team, os.delete_team, os.add_team_member, os.remove_team_member, os.return_file, os.add_work_item, os.update_work_item, os.close_work_item, os.claim_work_item, os.write_work_item_draft, os.publish_role, os.publish_team, os.set_steps, os.attach_repair, os.connect_tool, <prefix>.click, <prefix>.type, <prefix>.upload, <prefix>.select, <prefix>.press, os.use_tool.
A few reads have a side effect, listed on the tool:
os.mail_read: Marks the message read for the member, the same as opening it in the portal.os.my_mail_read: Marks the message read for you, the same as opening it.os.chat_read: Over the API and the MCP server, marks the thread read, the same as opening it in the portal. In a run, nothing.
Gates and grants
A task can put a gate on a tool it declares, in its gates, and the call then waits for the CEO. A gate on a workspace.* or memory.* tool waits like any other gated call. A memory.* call is checked against the memory grants first, so one they refuse fails without asking anyone. While a task’s gates cover workspace.git_push, the OS pushes none of the run’s work on its behalf, not even a checkpoint. Some built-in tools carry gates or grants of their own:
os.ask_ceois always gated, whatever the task declares, and the run ends on the CEO’s decision.os.mail_replyis always gated, whatever the task declares, and the run ends on the CEO’s decision.os.mail_sendis always gated, whatever the task declares, and the run ends on the CEO’s decision.os.escalate_mailis always gated, whatever the task declares, and the run ends on the CEO’s decision.os.set_gate_summarymust be called beforeos.mail_reply, which is handed back until it is. A task that declares one must declare the other.memory.train:taskis always allowed.memberandroleonly on an original task, never a clone.teamandenterpriseonly when this version grants that layer and the source event id starts with the granted prefix. A cloned task may too, in its own enterprise's memory and never its publisher's, as a task the enterprise wrote itself may; only a layer withdrawn on the team's page of a team the running member sits on or leads stops it, withclone_memory_not_granted. Reserved event ids (team:andgate:) stay refused.
What else a tool needs (a repo input, an assigned member, a binding) is on its own entry.
Retry safety
A call can be sent twice: a response lost on the way back, a retried run. Each tool says what a repeat does:
Errors
A failed call comes back one of three ways:
| Handling | What happens |
|---|---|
Handed back (hand_back) |
Refused and handed back: nothing ran, and the model gets another turn |
Tool error (tool_error) |
A tool error: the model reads it as the result and carries on |
Run fails (fail) |
The run fails |
Any tool call can come back with these, before the tool itself is reached:
| Error | When | What happens |
|---|---|---|
tool_not_on_task |
The tool exists, but the task does not declare it. | The run fails |
unknown_tool_name |
No tool anywhere has that name: a typo. The model is told the task's closest tools. | Refused and handed back: nothing ran, and the model gets another turn |
unknown_tool_repeated |
Three unknown names were already handed back in this run, and it called another. | The run fails |
bare_tool_name |
The call named a whole vendor catalog (github) instead of one of its tools. |
Refused and handed back: nothing ran, and the model gets another turn |
tool_value_not_declared |
The task pinned a field (<tool>#<field>=<value>) and the call carried another value. |
Refused and handed back: nothing ran, and the model gets another turn |
tool_value_refused |
The same, where the task marked the field fatal (<tool>#!<field>=<value>). |
The run fails |
Each tool lists its own errors on its entry.