Skip to main content
A Mission is a versioned task for the Milian copilot. It runs unattended on demand, on a schedule, or after a workspace event — no browser tab needs to stay open. Each run has a visible status, credit usage, and its own Milian transcript on the existing Milian Missions page.

Watch the walkthrough

This illustrated guide uses example data and reviewed product screenshots. It explains setup; it does not execute calls or actions. The active Mission version fixes the trigger, filters, referenced resources, connected apps, allowed action groups, and run limits. Milian receives only that published scope during a run.

Start from a template

When your workspace has no missions yet, Milian Missions (sidebar: Missions) opens with a field and a template gallery. Type a job into What should Milian do for you regularly? and send it, or pick one of the Recommendations below. Filter them by All, Self-improvement, Monitoring, Reports or Customers. Your own mission opens Milian with the setup prompt from New mission. Once you have missions, the same gallery opens under New mission → From a template; New mission → Describe to Milian starts without a template. Picking a template opens Milian with it. Milian asks only for what is still missing, such as which assistant or where the report should go, and then shows the usual confirmation card. Nothing is created or run before you confirm. Times use your workspace time zone: tell Milian another time during setup, or change it later under Trigger in Mission Studio.

Let your assistant improve itself

Assistant improves itself comes first and is marked Recommended in the template window. Every day at 07:00, it reviews yesterday’s calls and chats of one assistant with their evaluations: unanswered questions, wrong or missing information, and conversations that ended early or were rated badly. Then it sends a short report with links to the conversations, or one line if nothing needs fixing. Before the mission is created, Milian asks Fix it yourself or only suggest? and sets one of two modes:
  • Fix: Milian adds missing answers to the assistant’s knowledge base and sharpens its prompt where the prompt caused the problem. Prompt changes are saved as an assistant version, so you can restore an earlier one under Version history. The report lists every knowledge base entry Milian added.
  • Suggest: Milian changes nothing and lists its suggestions in the report.

Templates that run on an event (Beta)

Templates marked Beta run after every call or for every new contact. They appear only when Beta Features is on under Settings → Workspace. Milian saves them as a draft, and a workspace owner or admin publishes it in Mission Studio.

List templates through the API

The same list is available read-only through GET /api/v1/mission-templates (scope routines:read, optional language and category), the MCP tool list_mission_templates, the CLI command famulor list-mission-templates and the TypeScript SDK (client.api.listMissionTemplates()). Each template contains its prompt with placeholders such as {assistant}, the question behind each placeholder and the schedule fields for create_routine. Replace the placeholders before you create the mission. Event templates have no schedule fields and can’t be created through the API; set them up with Milian.

Mission Studio (Beta)

Open a Mission on /routines to use the Studio tabs: Overview, Trigger, Access, Runs, and Versions. Existing schedule-only Missions continue to run unchanged. Their first Studio edit creates a version-1 draft. Overview holds the status, name, instructions, referenced resources and thinking level; Trigger decides when the Mission runs; Access sets the allowed actions, connected apps and run limits. One published version has exactly one trigger. The Studio supports manual and scheduled runs as well as call, conversation, message, email, booking, lead, inbound-webhook, CRM, and connected-app events. An AND/OR filter group can narrow the event by safe event fields. Event Missions are a Workspace Beta feature. Turning Beta Features off pauses event Missions and marks them as needing attention; classic manual and scheduled Missions stay active.

Skip empty automation alerts

For a manual or scheduled alert, attach one to six automations in Overview, allow Read, and enable No-failures check in Access. Choose a lookback of 1–1,440 minutes and publish the version. Each run checks those automations for failures in that window. If every check succeeds and finds no failures, the run records its evidence and finishes without a model call or notification. An unreadable automation or incomplete check never proves there were no failures. The checks use the existing tool-call and runtime limits: allow at least one call per attached automation plus two control calls. If failures exist, Milian continues the objective with the remaining budget. The condition applies only after you explicitly publish it; other Missions keep their behavior. The version API accepts the optional readCheck property in the complete Mission specification: {"kind":"automation_failures","lookbackMinutes":60}. Save that complete specification through POST /api/v1/routines/{id}/versions or MCP save_routine_draft, review it, then use the existing publish operation. The CLI equivalents are famulor create-routine-version-draft <id> --spec @mission-spec.json and famulor publish-routine-version <id> --version-id <version-id>. The generated TypeScript SDK exposes client.api.createRoutineVersionDraft({path:{id},body:{spec}}) and client.api.publishRoutineVersion({path:{id},body:{version_id}}). Publishing still requires an owner or admin.

Schedule types

All schedules run against a timezone (default your workspace timezone) — the fields above are wall-clock time in that zone, including through daylight-saving transitions. A Custom cron schedule must fire at least 10 minutes apart; a tighter interval is rejected when you save it. An Hourly mission fires once per wall-clock hour at its chosen minute; when a daylight-saving transition skips or repeats a wall-clock time, the mission resolves to the first valid occurrence.

Required connections

Mission Studio grants individual connections already saved under Automations → Connections. A Mission refuses to run when a granted connection is missing, expired, disabled, or broken. Connections are rechecked immediately before each run.

Inbound webhook secret

For an inbound-webhook Mission, a workspace owner or admin generates the delivery secret in Mission Studio. The browser creates a random 32-byte value and shows it only once; copy it immediately and send it in the X-Automation-Secret header. Later visits show only whether a secret is configured. Rotate secret replaces it immediately, so the previous value stops working. Neither REST nor MCP ever returns a saved secret or the internal managed Automation identity.

Creating a mission

For prompts you can copy, see Milian Mission examples. While your workspace has no missions, you can also start from a template. Milian Missions are chat-first: ask Milian directly and describe the goal, trigger, filters, resources, apps, and allowed actions. New mission opens this flow. Milian prepares a review, and you can continue in Mission Studio before activation.
Milian composer with the New mission setup prompt

New mission opens Milian with a setup prompt. Send it to begin defining the mission, then review the proposed configuration before activation.

In both Milian composers, type @ to attach workspace context. Results include actual assistants, Missions, calls, contacts, and other workspace resources, plus only saved app connections and active Agent Tools. It never expands into the full app catalog. Assistant references show their avatar or the same orbit used on the Assistants page; phone numbers use their country flag; saved connections and Agent Tools use their vendor logo. Type / for commands that start a Milian review flow rather than mutating data silently. Creating, updating, or deleting a classic Mission through chat still shows an Approve / Cancel confirmation card. Publishing a Studio draft additionally shows the immutable version review before activation.

Managing a mission

The existing Milian Missions page lists trigger, status, active version, last run, next execution, credits, and attention state. Opening a Mission shows its Studio instead of a second page. Members may save drafts; workspace owners and admins may publish or pause. Viewer and Billing roles are read-only. Run now starts an off-schedule run; the transcript remains available from the row menu. A finished run can be retried manually: this always creates a separately marked run linked to the source run and never enables automatic retries. Publishing creates an immutable version. Later edits stay in a draft while the active version keeps running. Restoring an older version creates a new draft; it never rewrites history.

Background execution & transcripts

Every trigger creates a visible durable background run. Event runs are deduplicated and processed one at a time per Mission; additional events wait in FIFO order. The run appears as a Milian thread with the tools used. Run transcripts are kept for 10 days. Each mission also tracks its last run’s outcome (succeeded, failed, or still running) and the next scheduled run, both shown on the Milian Missions page. A run fails visibly when its owner loses access, the workspace is suspended, the plan or credits are unavailable, a connection was revoked, a capability boundary is crossed, or a run limit is reached. A possibly completed external mutation is never automatically repeated. Start a fresh manual run after reviewing the transcript. Operational telemetry covers publications, trigger and run outcomes, queue and runtime duration, credits, tool-failure counts, and safe failure categories. It never includes prompts, secrets, raw event payloads, or raw tool output.

Mission billing

Milian Missions capacity is sold as a monthly add-on unless included in the workspace plan. Every run uses normal Milian credits. A published version also sets maximum credits, tool calls, and runtime. Actual provider usage is always charged; a run that crosses its credit ceiling is marked failed and does not continue. A version can also set a thinking level for every run (the optional effort field of the version spec: low, medium, high or xhigh, as in the Milian composer); higher levels reason longer and use more credits, and without one the platform default applies.

Mission plan limits

The number of missions a workspace can have is its included plan capacity plus purchased Milian Missions units (-1 = unlimited, 0 = unavailable without the add-on). Creating a mission past that capacity is rejected until you delete one or add capacity.

API & MCP

Everything above is available in the public REST API and as MCP tools at https://<your-domain>/mcp: User-bound API keys and OAuth tokens keep that user’s workspace identity. Service-account credentials without a user principal fall back to the workspace’s oldest owner member.

What a mission may do on its own

A mission runs unattended, so its published version is the approval boundary. Grant only the action groups it needs: Read, Write, Communicate, and Call. A concrete tool call is checked again against that group and against its individual connection grant at the final execution boundary. Billing and plan changes, roles and membership, credentials and secrets, platform administration, destructive deletion, and subagent delegation stay off-limits during the Beta. Be deliberate with missions that read content you do not control — inbound email, call transcripts, web pages, records from a connected app. Instructions hidden in that text can influence the run. Give such missions a narrow prompt and review their transcripts. A Mission with the Communicate or Call grant can reach real people. These are real actions with real effects, so keep event filters and grants as narrow as possible and review run transcripts regularly. With Communicate, a Mission can send an email through the workspace email channel. Specify the recipient and report content in the objective. Sending still requires an available sender and sufficient credits. An unsuccessful send appears as a failed tool in the transcript; a preview does not send an email.