ai-telecom

Python SDK for AI Telecom — place AI phone calls with a number and a prompt.

mtds.vn · Platform · API reference · OpenAPI

---

pip install ai-telecom

from ai_telecom import Client

client = Client(api_key="sk_live_...")

call = client.calls.create(
    to="+84901234567",
    prompt="""
    Bạn là nhân viên tư vấn của MTDS.
    Chào khách, hỏi xem anh chị có quan tâm gói internet không.
    """,
)

print(call.id, call.status)   # call_Ku3n8pQ7... queued

That is the whole getting-started. No campaign, no contact list, no workflow to build first — a phone number and a prompt is the complete API surface you need.

base_url defaults to https://telecom.mtds.vn; pass it only when developing against a local deployment.

---

Calls

# Prompt inline — omit from_number → platform uses your first active SIP line
client.calls.create(to="+84901234567", prompt="...")

# Pick which number to call from (display number, SIP extension, or trunk name)
client.calls.create(
    to="+84901234567",
    prompt="...",
    from_number="0592014235",
)

# Pick a voice (Vietnamese default on the platform is Giang)
client.calls.create(to="+84901234567", prompt="...", voice="s6W2NupNY6TykGJoDtWy")

# Reuse a saved agent
client.calls.create(to="+84901234567", agent_id="agent_9fQz...")

# Per-call script; keep the agent's voice / KB / config
client.calls.create(
    to="+84901234567",
    agent_id="agent_9fQz...",
    prompt="Hôm nay chỉ nói về gói mùa hè…",
)

# Skip scoring / detach the agent's knowledge base for this call
client.calls.create(
    to="+84901234567",
    agent_id="agent_9fQz...",
    evaluation="",
    knowledge_base_id="",
)

# Many numbers at once → returns a list
client.calls.create(to=["+84901234567", "+84912345678"], agent_id="agent_9fQz...")

# Attach your own IDs (scheduling belongs in your own system — call create when ready)
client.calls.create(
    to="+84901234567",
    prompt="...",
    metadata={"crm_id": "C-42"},
)

The prompt is resolved in this order, first match wins: inline prompt → inline agent → agent_id → your default agent.

Empty string "" on evaluation, knowledge_base, or knowledge_base_id means explicitly off (do not inherit from the agent). Omit the field to inherit.

Whatever is resolved gets snapshotted onto the call, so editing an agent later never rewrites a call that already ran.

Reading results

call = client.calls.get("call_Ku3n8pQ7...")

call.status        # completed | no_answer | failed | cancelled | queued | dialing | ringing | in_progress
call.duration      # seconds of actual conversation (ringing excluded)
call.transcript    # [TranscriptTurn(role="ai", text="..."), ...]
call.score         # 0–100, only when you passed `evaluation`
call.outcome       # potential | follow_up | not_interested
call.summary
call.usage.cost_usd      # None means "no pricing configured", NOT free
call.usage.latency_ms    # answer → first audio the customer hears

# Recording (signed URL, expires in 10 minutes)
client.calls.recording_url(call.id).url

Listing and filtering

client.calls.list(status="completed", score_min=70)
client.calls.list(status="dialing,ringing,in_progress")   # anything live
client.calls.list(to="0901234567")
client.calls.list(from_="2026-07-01", until="2026-07-31")

# Walk every page without managing offsets yourself
for call in client.calls.iter(outcome="potential"):
    print(call.to, call.score)

Cancelling

client.calls.cancel("call_Ku3n8pQ7...")   # drops from queue, or hangs up mid-call

Waiting for a result

call = client.calls.wait("call_Ku3n8pQ7...", timeout=600)
print(call.score, call.summary)

wait() polls, which is fine for scripts. In production prefer the call.completed webhook — a call can legitimately run for minutes.

---

Scoring a call

Pass evaluation and the platform scores the conversation after it ends:

client.calls.create(
    to="+84901234567",
    prompt="Bạn là nhân viên tư vấn...",
    evaluation="""
    Chấm 0–100:
    - Khách có quan tâm sản phẩm không
    - Khách có hỏi về giá không
    - Khách có đồng ý hẹn gặp không
    """,
)

Omit it and the call still runs — it simply returns no score, outcome or summary, and skips the scoring model entirely.

---

Agents (optional)

An agent is a saved prompt + evaluation + voice. Use one when you want to reuse the same script across many calls.

agent = client.agents.create(
    name="Fibre internet outreach",
    prompt="Bạn là nhân viên tư vấn của MTDS...",
    evaluation="Chấm 0–100 theo mức độ quan tâm.",
    is_default=True,     # now calls.create(to=...) works with no prompt at all
)

client.calls.create(to="+84901234567", agent_id=agent.id)

client.agents.list()
client.agents.update(agent.id, prompt="...")
client.agents.delete(agent.id)

---

Knowledge bases (optional)

A knowledge base is a collection of documents. Reference material the AI consults when the customer asks — a price list, an FAQ, product specs.

# Empty collection, then Markdown document
kb = client.knowledge_bases.create(name="Bảng giá 2026")
doc = client.knowledge_bases.documents.create(
    kb.id,
    name="Gói cước",
    content="Gói Cơ bản: hai trăm nghìn đồng một tháng...",
)

# …or upload a file into the collection (PDF, DOCX, PPTX, XLSX, HTML, TXT, MD)
doc = client.knowledge_bases.documents.upload(kb.id, "bang-gia.pdf")
kb = client.knowledge_bases.wait(kb.id)   # until processing_count == 0

# Shortcut: create + upload + wait
kb = client.knowledge_bases.create_with_upload("bang-gia.pdf", name="Bảng giá 2026")

if kb.failed_count:
    print("Một số tài liệu lỗi — kiểm tra documents")
elif kb.is_ready:
    for d in kb.documents or []:
        print(d.name, d.status, d.effective_char_count)
    full = client.knowledge_bases.documents.get(kb.id, kb.documents[0].id)
    print(full.content)
    for s in full.sections or []:
        print(s.level, s.heading, s.char_count, s.enabled)

File upload returns immediately with document status="processing". Poll the knowledge base with wait() (or get()) until processing_count == 0 before using it — otherwise calls.create returns knowledge_base_processing or knowledge_base_unavailable.

wait() also returns when some documents failed (password-protected PDF, unsupported format) — check failed_count / each document's error.

Attaching it

# To an agent…
client.agents.create(name="Sales", prompt="...", knowledge_base_id=kb.id)

# …or straight into one call
client.calls.create(to="+84901234567", prompt="...", knowledge_base_id=kb.id)
client.calls.create(to="+84901234567", prompt="...", knowledge_base="Giá: ...")

# Detach the agent's KB for one call
client.calls.create(to="+84901234567", agent_id=agent.id, knowledge_base_id="")

Size is a real cost

There is no retrieval step: the entire enabled content of the knowledge base goes into the system prompt of every call that uses it. estimated_tokens is therefore per-call cost, not a one-off. Disable documents or sections you don't need in the dashboard — effective_char_count shows what is actually sent.

Write numbers as words ("hai trăm nghìn", not "200000") — the text is spoken aloud, and digits are read unreliably.

Phone numbers

Read-only; numbers are provisioned by an administrator (SIP line ↔ trunk). They are not configured on the Agent — pass from_number per call, or omit it and the first active number on the account is used.

for number in client.phone_numbers.list():
    print(number.id, number.number, number.trunk, number.quota_remaining)
    # number.id      → SIP extension, e.g. "1001"
    # number.number  → caller ID customers see, e.g. "0592014235"

# Explicit from_number — accepts display number, extension, or trunk name
client.calls.create(
    to="+84901234567",
    prompt="...",
    from_number="0592014235",   # or "1001" or "trunk-name"
)

Each number carries its own concurrency limit, minute quota and calling hours. Minute quota uses the same rule as Usage: each answered call consumes its exact duration_sec / 60.0 (not rounded up). A call created when the line is out of minutes is rejected (429 quota_exceeded).

---

Usage

usage = client.usage.get()                      # last 30 days
usage = client.usage.get(from_="2026-07-01")

print(usage.total.calls, usage.total.minutes)
for day in usage.buckets:
    print(day.date, day.calls, day.cost_usd)

minutes are actual minutes: each call with talk time contributes its exact duration_sec / 60.0, not rounded up — matching phone-number SIP quota and the TTS portion of estimated cost_usd. Example: three 30-second calls → 1.5 minutes, not 3.

---

Webhooks

Set a webhook URL on your API key in the dashboard. Every finished call is POSTed to it, signed with HMAC-SHA256 in the X-Signature header.

from ai_telecom import webhooks
from ai_telecom.webhooks import WebhookVerificationError

@app.post("/hooks/ai-telecom")
def handle(request):
    try:
        event = webhooks.verify(
            request.body,                      # RAW bytes — do NOT json.loads first
            request.headers["X-Signature"],
            secret=WEBHOOK_SECRET,
        )
    except WebhookVerificationError:
        return 400

    call = event.data                          # a full Call object
    print(call.id, call.score, call.metadata)
    return 200

Verification is constant-time. Always verify before trusting a payload.

---

Errors

from ai_telecom import (
    AITelecomError,        # base class — catches everything below
    AuthenticationError,   # 401
    PermissionError_,      # 403 — key lacks the scope
    NotFoundError,         # 404
    ValidationError_,      # 400 / 422
    RateLimitError,        # 429 — has .retry_after
    ServerError,           # 5xx
    APIConnectionError,    # network failure
)

try:
    client.calls.create(to="not-a-number", prompt="...")
except ValidationError_ as e:
    print(e.code, e)       # invalid_phone  `not-a-number` is not a valid phone number

Network errors and 429/5xx are retried automatically with backoff. Logic errors (401/403/404/422) are never retried — repeating them cannot change the answer.

---

Configuration

client = Client(
    api_key="sk_live_...",
    base_url="http://localhost:3000",   # local development
    timeout=30.0,
    max_retries=2,
)

with Client(api_key="sk_live_...") as client:   # closes the HTTP pool on exit
    ...

---

Versioning

This SDK is pinned to /api/v1. A breaking server change ships as /api/v2 alongside a major SDK version — v1 is never changed underneath an installed client. Unknown response fields are ignored rather than raising, so an additive server change never breaks an older SDK.

License

MIT © MTDS