Decision brief · October 2026

Let the agent make the call—not translate one you join

$18.54100 cloud minutes · landlines
AI + carrier + one foreign number

$40.28100 cloud minutes · mobiles
AI + carrier + one foreign number

$10.78 + bridge100 local-SIM minutes · AI only
SIM tariff and hardware additional

Published-tariff budgets, before tax. No live call tested. Local-SIM feasibility depends on the exact phone and a working two-way audio bridge.

Research checked October 6, 2026. Scope: fully autonomous Turkish calls for practical administrative tasks. No live calls, paid accounts, purchases, rooting, or phone changes were performed. This is a feasibility and budget report, not a working deployment.

The fastest documented route is Retell plus Twilio; the existing SIM deserves a bounded hardware test

Recommendation: use Retell as the conversational layer. Test the existing Android as a local cellular gateway before buying new phone hardware, but do not assume a USB cable solves the audio connection. If that test is awkward or unreliable, Retell plus an independently owned Twilio SIP trunk is the clearest documented cloud route into Turkey. Retell documents Turkish speech recognition and speech synthesis, custom telephony, and Twilio SIP integration.⁵⁶¹³

The intended workflow is simple: send the agent a task in English; it calls in Turkish, asks the bounded questions, and returns the outcome in English. No joining the call, no live interpretation, no transfer to the user. If a business refuses to deal with an assistant, the agent stops and reports that blocker rather than pretending to be the account holder.

An Android can supply the cellular line, but call control and call audio are different problems

A usable gateway must handle dialing, answer detection, DTMF, hangup, hearing the remote person, and sending generated speech into the cellular microphone path. Success at any one of those does not establish the rest.

Ordinary Android apps cannot simply request cellular call audio: the official VOICE_CALL, VOICE_UPLINK and VOICE_DOWNLINK capture sources require CAPTURE_AUDIO_OUTPUT, reserved for system components.² scrcpy now exposes call-audio source flags, so saying it can never capture calls would be wrong; their existence still does not prove capture works on a particular phone, and recording the uplink is not injecting speech into it.³

The most defensible Mac-compatible experiment uses the phone's headset interface

Proposed path: Android + Turkish SIM ↔ full-duplex headset interface ↔ USB audio interface on the Mac ↔ real-time Turkish voice agent. A separate control connection handles dial/hangup/DTMF. This is an integration design, not an already tested chain.

There is established hardware for the audio portion. JK Audio's Daptor Two manual explicitly says it can send and receive audio at the same time during a regular cellular call and that the phone recognizes it as a headset.¹ This establishes a physical route around the need for an app to intercept cellular audio. It does not establish compatibility with the unidentified phone, unattended dialing, or an AI bridge.

A phone without a headset jack needs a compatible USB-C adapter with microphone support, not a headphone-output-only dongle. The exact adapter, concurrent charging, control connection, gain levels and echo isolation require a device-level test. The manufacturer also cautions that not every phone is guaranteed compatible.¹ A Mac USB audio interface and a real-time software bridge are additional requirements. Do not buy the named interface yet: current Turkey stock, delivered price and the complete bill of materials are unverified.

Other local-SIM paths have material limits

Bluetooth HFP through Asterisk is a real architecture: the upstream mobile channel contains call control and SCO audio handling.¹⁵ A Linux Bluetooth bridge is a candidate, not proof that ordinary Android–Mac pairing or Docker Desktop supplies full-duplex call audio. Exact current phone/adapter compatibility remains untested.
Rooted Android SIP gateway apps are not the default recommendation. Zerkat advertises root as mandatory, Samsung as officially supported, and manual server authorization for each device.¹⁴ Its claims are not an independently verified stable service for this phone. Rooting or unlocking is outside this research authorization.
GSMCalls advertises no-root phone gateways, but its compatibility page explicitly says its rows are design-preview placeholders before launch.¹⁰ The published minimum is $100/month for up to ten phones, before carrier airtime and hardware.¹¹ Reject it as a validated one-phone recommendation; a polished compatibility table is not a lab result.
Netgsm IP FCT is a commercial SIM/eSIM-to-SIP alternative, advertised at 810 TL/month per port. It supports SIP-trunk PBXs and one simultaneous call per GSM port; incoming calls are busy during use.⁹ Existing-operator compatibility, personal foreign-resident eligibility, any required PBX subscription, tax treatment and setup costs remain unresolved. That is a quote candidate, not a verified 810 TL all-in solution.

The screenshot identifies the phone as G316, which matches the General Mobile GM 23. The maker confirms Helio G37, Bluetooth 5.0, VoLTE, USB-C and a 3.5 mm headset/microphone jack; its published software spec says Android 12, while your screenshot shows Android 13 installed.²³²⁴ That is enough to narrow the prototype: start with stock-phone Bluetooth HFP, and retain the 3.5 mm jack as a wired fallback. It still does not prove this handset routes cellular audio to HFP/headset in the way the bridge needs.

The existing SIM is Turkcell, and the phone can stay beside the Hermes Mac, powered. That makes the local-SIM route the first one to test; no virtual +90 number or cloud trunk needed for the prototype.

The strongest documented route is stock Android → Bluetooth HFP → small Linux appliance → Hermes Mac. Linux handles call control and separate send/receive audio; the Mac runs the voice agent. Published engineering notes document dialing/hangup and simultaneous audio capture/injection through oFono and PipeWire.¹⁸¹⁹²⁰ That proves the architecture, not compatibility with this handset.

Planning estimate: $70–180 one-time hardware if a small Linux appliance and compatible Bluetooth adapter must be bought; less if suitable Linux hardware is already available. A Mac-only wired-audio design is possible but more cumbersome, estimated at $250–450 for interface, adapter and cables. These are estimates, not Turkey quotes. AI usage remains about $10.78 per 100 connected minutes; add your actual Turkcell plan charge, which we have not verified.

Keep the SIM in the phone for now. Turkcell’s terms exclude FCT devices on at least some named tariffs, and its M2M notice covers relevant M2M/FCT lines.²¹²² This does not prove your ordinary consumer tariff is barred. We need the exact plan before putting the SIM into any separate gateway. No call, purchase, phone change or installation has happened.

The AI costs about $0.11/minute in this explicit baseline

Modeled configuration: Retell voice infrastructure $0.055/minute + ElevenLabs Flash $0.040/minute + GPT-4.1 mini $0.0128/minute = $0.1078/minute.⁴ Use a Turkish-capable Flash v2.5 voice and Turkish recognition; Flash v2 is English-only.⁵ This is a cost baseline, not a claim that this model already handles Turkish negotiations reliably.

Retell's custom telephony fee is zero, but the carrier, any gateway service, and bridge hosting are separate.⁴ A local-SIM implementation would therefore have an AI subtotal of $10.78 at 100 minutes or $32.34 at 300 minutes, plus its actual SIM plan and the bridge. If its domestic minutes already cover the usage, incremental carrier airtime could be zero within the allowance; the plan and allowed usage have not been verified.

For the cloud route, Twilio Elastic SIP termination is $0.0661/minute for ordinary Turkey landlines and $0.2835/minute for mobiles. A foreign number starts at $1.15/month.⁷ Do not add a separate Programmable Voice destination charge or Media Streams fee to this SIP calculation.

Costs for five-minute connected calls

One call: local-SIM AI subtotal $0.54; cloud landline usage $0.87; cloud mobile usage $1.96. Cloud number rental is separate from these per-call figures.
Twenty calls / 100 minutes per month: local-SIM AI subtotal $10.78; cloud all-landline total $18.54; cloud all-mobile total $40.28.
Sixty calls / 300 minutes per month: local-SIM AI subtotal $32.34; cloud all-landline total $53.32; cloud all-mobile total $118.54.

Monthly cloud totals include one $1.15 foreign number. Local-SIM subtotals exclude hardware, bridge software/hosting, carrier charges and any gateway subscription. All USD totals exclude taxes, optional add-ons, special destinations, retries, long-prompt adjustments and setup work. Calls are modeled at exactly five connected minutes. Holding and silence count toward Retell billing, and connected voicemail is billable.⁴ Twenty short calls are not necessarily 100 billable minutes if a call center leaves the agent on hold.

Cloud formula: $1.15 + connected minutes × ($0.1078 AI + $0.0661 landline or $0.2835 mobile). Calculations were run in Python using decimal arithmetic.

One-time setup is still work, not a vendor setup-fee quote

Cloud setup needs personal Retell/Twilio accounts, funding, number provisioning, Turkey geographic permissions, Turkish agent configuration and a Hermes task/result adapter.⁶ Exact account approval and initial cash deposits have not been checked in checkout. No additional server purchase is established as necessary for this route; existing orchestration infrastructure can be evaluated during implementation.

The local-SIM route additionally needs a phone-compatible audio/control bridge. No verified one-time hardware total is available yet. Do not describe it as free just because the phone and SIM already exist.

The cloud trade-off is a foreign caller ID

Twilio explicitly permits international calls into Turkey but prohibits Turkish or anonymous caller ID on that route.⁸ Verifying an existing +90 number does not override that restriction. Recipients may screen foreign numbers; no answer-rate measurement is available.

A genuine domestic cellular call uses the existing SIM rather than asking an international carrier to present a Turkish identity. That is why the local-SIM route is worth evaluating. Caller-ID behavior, SIM terms and real call completion still need testing. Do not buy a Turkish virtual number and assume that alone creates compliant domestic routing.

Alternatives do not remove the audio or carrier problem

ElevenAgents is a credible alternative voice layer: current public hosting plans are $6/month for 75 minutes, $22/month for 275 minutes, and $0.08/additional minute; LLM and telephony are separate.¹² It does not magically expose an Android SIM's audio. Its exact all-in cost needs a selected LLM and measured usage, so no false exact comparison is given here.

Netgsm domestic fixed/SIP service lists 1.10 TL per 60-second block to ordinary domestic fixed/mobile destinations, including its stated taxes.¹⁷ That is not the tariff of an existing mobile SIM. Personal eligibility and end-to-end AI integration were not established, so it remains a separately scoped option.

Hermes already publishes an optional telephony helper for Twilio/Vapi/Bland outbound calls.¹⁶ Local read-only diagnosis found no configured Twilio or Vapi calling setup. Inspecting the helper showed that its default Vapi payload does not explicitly configure Turkish recognition or DTMF tools. Existing helper code is a starting point, not proof of a ready Turkish calling service.

The first test must prove an actual conversation, not just a ringing phone

Recommended acceptance test, after setup/spend authorization:

Use a consenting test endpoint, not a random business. Confirm the displayed caller ID and inspect the actual connected duration and charge.
Prove independent audio both ways, including speaking over the agent, background noise and no self-echo loop.
Test Turkish names, amounts, appointment dates, spelling and corrections. Repeat prices and commitments back before treating them as confirmed.
Exercise a keypad menu, hold, remote hangup and maximum-duration termination.
Return a concise English result with explicit answered/unanswered questions. A completed carrier call is not a completed task.
Enforce no unauthorized charges, repairs, legal/medical commitments or account authentication. Identify the caller as an assistant. Stop on refusal. Configure recording/transcript retention explicitly; default to minimal retained data.

The remaining hardware-selection input is the Android model and version. Turkcell is identified and the phone can stay beside the Mac, but the exact plan still needs confirmation before any SIM-as-gateway route beyond using it in the phone. Until the model is known, autonomous Turkish calling has a documented cloud path and a credible local-SIM architecture, but the local deployment remains untested.