phone-call

name	phone-call
description	Place outbound AI phone calls via Bland AI: book reservations, make appointments, ask businesses questions, then report back the transcript and summary.
license	MIT
allowed-tools	skill_run read_skill
metadata	{"gini":{"version":"1.0.0","author":"Gini","platforms":["macos","linux"],"requires":{"credentials":["BLAND_API_KEY"],"approval":["place-call"]}}}

Phone Call

Place outbound phone calls through Bland AI. An AI voice agent dials the number, follows your task prompt in a live conversation, and the full transcript plus a summary come back when the call ends. Everything runs through skill_run:

skill_run({ skill: "phone-call", script: "place-call",
            args: { phoneNumber: "[phone]", task: "..." } })

The runtime injects BLAND_API_KEY into the scripts — you never see or pass the key.

Workflow

1. Gather the full task. Before anything else, collect: who to call (name + number), the goal, hard constraints (dates, times, party size, budget), fallbacks if the first ask isn't available, and the name to give if the callee asks who's calling. A vague task produces a bad call.
2. Announce, then dial — the approval card is the confirmation. In one short message, state who you're calling (the exact number) and what the agent will say or ask, then call place-call immediately. Do not ask "should I go ahead?" in chat: place-call always pauses on an Approve/Deny card, and approving that card is the user's go-ahead. A denied card ends the task; if the user sends a new message with changes, adjust per their feedback and submit a fresh place-call — never re-submit the identical call.
3. Place the call with place-call. It returns { ok, callId }.
4. Wait for the result with check-call, passing waitSeconds: 240 — the script polls Bland every 10 seconds internally and returns as soon as the call completes (or when the budget runs out). If the result comes back with completed: false, call check-call again with the same args and repeat until completed is true.
5. Report back the summary and the key points of the transcript (quote relevant exchanges, don't dump the whole thing unless asked). Optional: for structured answers about the call (did they confirm? what time?), run analyze-call.

If the user wants to abort a call in progress, run stop-call with the callId.

Scripts

place-call

skill_run({ skill: "phone-call", script: "place-call", args: {
  phoneNumber: "[phone]",   // required, E.164
  task: "...",                   // required, the call prompt
  voice: "maya",                 // optional, Bland voice id/name
  firstSentence: "...",          // optional, exact opening line
  waitForGreeting: true,         // default true — wait for the callee to speak first
  record: false,                 // default false — see Rules
  maxDurationMinutes: 10,        // default 10
  language: "en-US"              // optional
} })

Returns { ok, callId } or { ok: false, error }.

check-call

skill_run({ skill: "phone-call", script: "check-call", args: { callId: "...", waitSeconds: 240 } })

Returns { ok, callId, status, completed, answeredBy, callLengthMinutes, to, from, transcript, summary, recordingUrl, errorMessage }. waitSeconds (optional, default 0, max 240) makes the script wait for the call to finish, polling Bland every 10 seconds; if the budget runs out first it returns the latest status with completed: false — call again with the same args. Fields Bland hasn't populated yet are omitted; transcript and summary appear only after completed is true. callLengthMinutes is in minutes. answeredBy distinguishes human from voicemail.

stop-call

skill_run({ skill: "phone-call", script: "stop-call", args: { callId: "..." } })

Ends an in-progress call. Returns { ok, message }.

analyze-call

skill_run({ skill: "phone-call", script: "analyze-call", args: {
  callId: "...",                                      // required
  goal: "Book a dinner reservation",                  // optional context
  questions: [                                        // required, non-empty
    ["Did they confirm the reservation?", "boolean"],
    ["What time was booked?", "string"]
  ]
} })

Use after check-call reports completed: true when you need structured answers instead of reading the transcript — e.g. "Did they confirm the reservation?" → true. Each question is a [question, answerType] pair ("string", "boolean", "number"); a bare string question defaults to "string". Returns { ok, answers } with one answer per question, in order. Each analysis costs Bland credits, so prefer the transcript/summary for simple cases.

Background watching

The workflow above is synchronous — right for short errand calls where the user is waiting on the result. When the user doesn't want to wait, or the call may run long, hand the waiting to a scheduled job with a call-watch pre-run hook: the hook polls Bland with zero model turns while the call is in progress, and wakes the job's turn exactly once, when the call finishes.

1. Place the call with place-call as usual and note the callId.
2. Create the watcher job:

create_job({
  name: "call-watch <callId>",
  intervalSeconds: 30,
  oneShot: false,        // must stay false — the silent in-progress ticks would auto-pause a one-shot job before the call finishes
  timeoutSeconds: 120,
  preRunHook: {
    handlerId: "skill-script",
    config: { skill: "phone-call", script: "call-watch", callId: "<callId>" }
  },
  prompt: "A phone call placed in the background has finished. The call result (status, summary, transcript) is in the fenced context items above — report the summary and the key transcript exchanges. Then find this job by name with list_jobs and delete_job it; the call is done and the watcher is no longer needed."
})

3. Tell the user the call is underway and the result will arrive in the job's chat thread, then end your turn — do not poll with check-call.

Notes:

• The report lands in the job's own chat thread (scheduled jobs deliberately do not post into the main chat).
• call-watch is the job's hook script, not for direct skill_run use. A failed call (never answered, rejected) also wakes the turn, so the user always hears the outcome.

Writing a good task prompt

The task is the agent's script for the whole conversation. Include the goal, context the callee will ask about, constraints, and fallbacks. Example:

You are calling Luigi's Restaurant to book a dinner reservation. Book a table
for 4 people this Friday at 7:00 PM under the name Shelden. If 7:00 PM is not
available, try 7:30 PM, then 6:30 PM. If none of those work, ask what times
are available Friday evening and say you'll call back. If they ask for a
phone number, give [phone]. Be polite and concise.

Use firstSentence when the opening line matters (e.g. "Hi, I'm calling on behalf of Shelden to ask about your store hours.").

Limitations

• Outbound only — this skill cannot answer incoming calls.
• Rate limit: about 1 call per 10 seconds to the same number.
• Transcript and summary are only available after the call completes — check-call during the call shows status only.

Rules

1. Never ask for confirmation in chat — the Approve/Deny card on place-call is the confirmation. State the number and task in your message, call place-call, and let the user decide on the card. Calls reach real people and cannot be un-placed. A denial ends the task; if the user follows up with changes, submit a fresh place-call reflecting them — never re-submit the same unchanged call.
2. Never call emergency numbers (911, 112, 999, etc.) under any circumstances.
3. Phone numbers must be E.164 format: [phone].
4. Don't store third-party phone numbers in memory — only the user's own.
5. Leave record: false unless the user explicitly asks for a recording — recording-consent laws vary by region.
6. Never include BLAND_API_KEY in a reply or tool argument — the runtime injects it into the scripts.