MCP Workflows

A tool list isn't a plan. Here are the patterns worth copying. Each is an instruction you give the agent, followed by the tools it calls.

1. Fleet-wide engagement with a saved prompt

Run my "Warmup — TikTok" prompt on every free TikTok device.

The agent calls list_devices, keeps the free TikTok devices, then calls start_run with prompt_id and those pad_ids. You get a group_id and a session link.

Saved prompts are the most repeatable option, because they carry persona, brand, model, and step limits. Reach for them whenever the work repeats.

2. Preview before you commit

Preview a lurk engagement for these five pad IDs.

The agent calls preview_run. It returns, per device, which template and persona will be applied. Nothing starts. This is the cheapest way to catch a misconfigured profile before it acts.

3. Profile-driven engagement

Start a recovery engagement on my Instagram devices, but only the free ones.

The agent calls list_devices, then start_run with engagement="recovery" and platform="instagram". Each device applies its own prompt and persona. Unless you pass max_steps, the midpoint of the recommended band is used (100–150 for recovery).

4. Post one video to multiple platforms

Ingest this Drive link and post it to TikTok and Instagram, capped at 200 steps.

Two calls:

1. ingest_content(url=...) → conbersa-file:// URI.
2. start_content_run(groups=[tiktok, instagram], files=[...], max_steps=200).

Media is pushed once; platforms run in sequence. Open the returned session link to watch. (Content tools are experimental.)

5. Status without blocking

Is that run still going?

get_session_status(group_id) returns running, completed, failed, or stopped, with per-device detail. Agents should poll this on demand rather than hold a stream open. The video stays in the web cockpit behind the session link.

6. Schedule recurring distribution

Schedule a full engagement on these five TikTok devices every day at 9am Eastern.

create_schedule with run_type="agentic_group", engagement="full", schedule_type recurring, timing_mode="specific", timezone="America/New_York", and scheduled_at. If you want the timing to look human, use timing_mode="random" with random_window_start, random_window_end, and a max_runs_per_day cap.

7. Find and fix a running schedule

List my schedules, then stop the one named "Nightly warmup".

list_schedules → stop_schedule(schedule_id). The agent can also run_schedule_now to fire one on demand, or edit_schedule to change the timing without rebuilding it.

Composing workflows

Chain read → preview → run → status:

1. list_devices — find free pads.
2. preview_run — confirm behavior.
3. start_run — execute.
4. get_session_status — report progress.
5. stop_run — step in if needed.

Once the chain is reliable, wrap it in create_schedule so it repeats on its own.

Guardrails worth setting

• One device for the first run of any new task.
• Use saved prompts for anything recurring; free-form tasks drift.
• Cap steps. max_steps keeps a stuck agent from running long.
• Filter busy devices before starting, to avoid lock errors.

Next

• Tool reference — arguments for every tool.
• Posting to multiple accounts — the content path in depth.