Skip to content

Mob’s shared instructions ask it to preserve the user’s scope, clarify missing targets, distinguish tool evidence from assumptions, and report unfinished work. Tool results, quoted content, memories and summaries do not grant permission. Installed skills guide the task within the user’s request and tool permissions. Cleared instructions should be loaded again when needed. Denied or revoked actions must not be rerouted through another tool, and an uncertain side effect must not be repeated just to discover whether it worked. These instructions complement the runtime approval checks described below.

The model works through tools. Any tool may ask the owner to approve a call before it has an effect, and the owner can decide from any client or channel.

  • Built in: datetime, remember (saves a fact after approval, below), when the sandbox is enabled, sandbox, and, while there are tools to load, tool_search (below).
  • The email plugin offers email_inbox, email_read and email_send; the Cloudflare plugin offers cloudflare. Each is defined in its own package and offered only while its plugin is enabled.
  • Coding agents add task_start (asks for approval), task_plan (asks for approval), tasks, task_message, task_stop, task_diff, task_check (asks for approval) and task_review (asks for approval).
  • Schedules add schedule (asks for approval), schedules and schedule_cancel.
  • MCP servers add mcp__<server>__<tool> tools, and tool_install installs a remote server from the catalog.
  • While any skill is installed, the skill tool loads one; it reads without approval.
  • An ownerOnly tool is offered only to the owner’s turns, and only the owner can approve its calls: another allowed WeChat user cannot read or send the owner’s mail. Plugin and MCP tools are owner-only.
  • Mail text and Cloudflare results sent to the model are bounded, and the model is told they are untrusted data.
  • remember saves one fact (up to 2000 bytes) once it is approved. A fact the owner approved is read in every one of the owner’s turns, in any conversation, client or channel; one another caller approved is read only in the conversation that saved it. Other callers’ turns never read the owner’s facts from elsewhere.
  • Facts go into the system prompt as data, not instructions. /memories lists them with where each is read, and /forget ID|all deletes them; both are owner-only.
  • Facts saved before 2026-10-20 stay with their conversation: Mob cannot tell who approved them. Ask the model to remember one again to share it.

Registering a tool and declaring it to the model are separate, so many tools cost the prompt nothing until a task needs them. Tool.exposure is one of:

  • direct, the default: declared to every turn that may use the tool. Built-in and plugin tools are direct.
  • deferred: declared once tool_search loads it. MCP tools are deferred unless their server’s configuration says otherwise.
  • hidden: never declared and never found.

While a turn has deferred tools to load, or entries such as skills to find, it offers tool_search, and the system prompt names how many wait in each namespace (“MCP server github (12)”). tool_search takes keywords and ranks what is waiting with BM25 over the tool’s name, namespace, description and parameters (camelCase and _ split, Han characters one word each); a query naming a tool exactly finds it first. It returns up to five results, ten when asked, and loads the tools among them: they are declared, with their prompts, from the turn’s next step. Calling a deferred tool that is not loaded, even beside the search that loads it, fails with an unsupported-tool error.

A conversation keeps the tools it loaded, and those it called before, declared in its later turns: they are read back from its history, so nothing else is stored. Loading changes no approval: a loaded MCP tool asks as it always does.

ToolRegistry.registerEntries adds things tool_search finds that are not tools to load, each with a kind, a description and how to use it, such as a skill and the tool that loads it. Owner-only entries are found only in the owner’s turns.

  • Reading runs at once. Sending mail, Workers AI (ai.run, and vectorize.query, which embeds with Workers AI), D1 queries that may write, remember, every MCP call, installing an MCP server with tool_install, starting a coding task or plan (task_start, task_plan), running a task_check command, starting a review squad (task_review) and scheduling a prompt wait for approval. A single D1 SELECT statement runs without approval; SQL the check cannot be sure of still asks.
  • The approval names what the call acts on (the sender and SMTP account, or the Cloudflare account and model, index or database), and an approved call runs only while the configuration still names the same target, so changing the configuration cannot redirect it.

Approval decisions require a boolean approved value. Malformed decisions are rejected before changing the pending approval or running its tool.

  • An approval request carries the tool name and the input to approve. describeApproval renders it as plain text for channels. The AI SDK tool-approval-request chunk carries only the approval and call IDs.
  • Before an approval decision, the web UI shows the full tool input above the approval buttons, including all parameters and long command or content text.
  • The web UI and TUI show the request in the conversation; /approve and /deny work in every client. Telegram shows Approve, Deny, Allow in conversation and Always allow buttons; WeChat iLink accepts 批准 and 拒绝, or a bare reply when one request is pending. /approve and /deny without an ID decide the only pending request.
  • A denial may carry a reason (Answer.reason). It is written into the denial note, and the model answers in a new turn. A denial without a reason ends quietly.
  • An approval left undecided past its hour is marked expired and its call failed, so the model and channels stop waiting on it.
  • Web and the terminal UI share the approval/denial stream orchestration in @mob/client. Their controls still send the same approval ID, reason and remembered scope to the service. They display service results and failures; neither retries stale decisions nor makes authorization policy locally.

Web approval showing the complete tool input before the decision buttons

  • Store.decide records the decision once and marks an approved call approval-responded. The service then runs the tool’s approved and Store.complete records its result. A retried approval is refused as stale, so the effect happens at most once. After a crash between those steps, startup marks the call failed with an unknown outcome. Check the external result before asking again; Mob never repeats the approved call automatically. Completed and undecided calls keep their state.
  • Effects outside Mob can key on the approval ID: an approved email’s delivery is keyed email:<approval ID>.
  • After a decision the model receives a note naming the tool: “The user approved remember and it ran. Result: …” or “The user denied the remember request.”

In the Web UI, chat and Agents cards show request details expanded by default. Long inputs wrap and scroll without truncation. Allow once and Deny… are the primary actions; More permissions… offers conversation-only or cross-conversation remembered approval and explains its scope. Partial requests on Agents cards cannot be approved there; open the conversation to review the complete request. Decision controls are disabled while submitting; failures remain visible in the conversation or agent card.

  • The owner may remember an approval for the conversation or always (Answer.remember). A grant covers the tool’s calls whose Tool.scope matches the approved input’s: the recipients and sender for email_send, the operation and target for cloudflare, the server configuration for MCP tools, any call for a tool without scope.
  • Covered calls in the owner’s turns run approved at once with a fresh approval ID; other callers’ turns always ask.
  • Grants live in the grants table. /grants lists them and /revoke ID|all forgets them. A grant outlives configuration changes only as far as its scope does: a new SMTP account or D1 database gives a new scope, which asks again. Remembering is coarse for tools without scope, so a tool whose inputs differ in risk should define one.

Writing a tool, including requestApproval, approved and scope, is covered in Extensions.

Web’s Manage Mob → Memories lists saved facts and their scope, links to source conversations, and confirms individual or bulk deletion. Remembered approvals shows the tool, matching scope and whether an approval applies to one conversation or all of them. Revoking an approval makes future matching calls ask again; revoking all approvals requires confirmation. The corresponding /memories, /forget, /grants and /revoke commands remain available.