Temsor API — Turkey & EU business data MCP Server

Notify an optional audio chunk / customer URI for a live or archived call (idle $0). Prefer customer storage URI. Temsor keeps a pointer; STT only while a call is active.

List recent calls with transfer path crumbs — site + customer archive browser. GET /v1/pbx/calls?limit=20. Scoped by x-temsor-pbx-key when set. Idle $0. No DID sales.

Get one call by id with transfer graph — customer detail drawer. GET /v1/pbx/calls/get?call_id=…. Scoped by x-temsor-pbx-key. found=false if missing. Idle $0.

Handoff graph + transfer edges — who→whom→where the promise stuck (moat). Idle $0. GET /v1/pbx/calls/graph?call_id=…&stuck_at=310. Returns path, path_tr (plain-language summary), edges, and raw transfers. Compounding asset for search/orchestration: handoff edges, not a bot transcript. Idle $0. No DID sales / no live PSTN.

Managed SaaS domain dictionary (E4) — example packs thicken per pilot (tenant-scoped aliases). Boosts search + Fact extract. No invented premiums. GET /v1/pbx/dictionary?domain=hotel|pbx_provider|insurance|finance|energy|automotive|all (legacy otel|santral_firmasi|sigorta|finans|enerji|otomotiv|msp still accepted). Idle $0, no LLM. Synonyms feed archive search (match crumbs) and transcript Fact extraction. Write path: POST /v1/pbx/dictionary/add.

Add a synonym/alias (tenant-scoped or global demo) — E4 compounding thickens search boost + Fact extract. ≤60/hour/key → 429. Idle $0. No invented premiums. POST /v1/pbx/dictionary/add {term, alias, domain:hotel|pbx_provider|insurance|finance|energy|automotive, scope?:pbx|global}. Managed SaaS compounding (E4): tenant-scoped aliases thicken search + fact extract. No invented premiums/prices. Idle $0, no LLM. Legacy keys (otel, santral_firmasi, msp…) still accepted.

Ingest call-ops events → open promises + transfer crumbs for search/orchestration. Optional HMAC. Idle $0 when no calls. Demo stores events in memory/file. Auth: omit, x-temsor-pbx-key, or x-temsor-signature = hex(HMAC-SHA256(api_key, body)); bad signature rejected when header present. BridgeEnter → transfer/handoff crumbs. Utterance/text → open Promise docs via heuristic (no LLM). Idle $0. Body ≤256 KiB → 429. Advanced: ARI/Stasis batch shape when customer runs own PBX — not Asterisk-first product copy.

List extracted facts (IMEI, room, tariff…) from the durable index — filter by call_id, kind, domain, q. GET /v1/pbx/facts?call_id=&kind=&domain=hotel|pbx_provider|insurance|finance|energy|automotive|all&q=&limit=50. Facts come from transcript extract + domain dictionary; persist in pbx-index. Scoped by x-temsor-pbx-key. Idle $0. No DID sales. Right-to-erasure ready. Advanced: own PBX/Asterisk stays on customer.

Get one extracted fact by id — detail drawer for site + customer. GET /v1/pbx/facts/get?id=…. Scoped by x-temsor-pbx-key. found=false if missing. Idle $0. No DID sales.

Upsert mid-call partial index (utterances → fact/promise hints). Path to ≤400ms whisper; ≤600/hour/key → 429; idle $0. POST /v1/pbx/index/partial {call_id, text?, speaker?, t_ms?, is_final?, live?, end_call?, promote?}. Streams partial STT queue into accumulating Fact/Promise hints while the call is live. No Temsor STT/LLM — customer posts text. Managed SaaS in-call index; no DID sales. Own PBX optional footnote. Idle $0.

GET mid-call partial index — accumulating facts/promises/snippets while live. GET /v1/pbx/index/partial/get?call_id=…. Returns live partial state for whisper ≤400ms path. Idle $0 when no live call. Managed SaaS in-call index; no DID sales. Own PBX optional footnote.

Heuristic promise extractor from transcript text — creates open Promise docs (idle $0, no LLM). Detects promise-like sentences (we will, I promise, refund, cancel, call you back, tomorrow…). Optional persist_promises=false for extract without writing. Also wired via utterance text on POST /v1/pbx/events.

Light CRM from call graph / facts — caller phones+names and extension identities seen on transfers. GET /v1/pbx/persons?call_id=&kind=caller|extension|all&q=&limit=50. Durable Person rows from pbx-index plus derived extension identities (ext_NNN) from Transfer edges / promise.ext. Scoped by x-temsor-pbx-key. Idle $0. No DID sales. Right-to-erasure ready. Advanced: own PBX/Asterisk stays on customer.

Get one light-CRM identity by id — caller Person or derived extension. GET /v1/pbx/persons/get?id=per_001|ext_101. Scoped by x-temsor-pbx-key. found=false if missing. Idle $0. No DID sales.

List open/broken call promises for agent handoff tracking — hotel / MSP sales demo. Idle $0. GET /v1/pbx/promises?status=open returns open/unclosed promises with transfer/handoff path for orchestration. Scoped by x-temsor-pbx-key when set. Idle $0. No live PSTN.

E1 moat: who→whom→promise → orchestra handoff intent (whisper/transfer/hold). Idle $0. No live PSTN. Own PBX = advanced footnote. POST /v1/pbx/promises/act {promise_id, action?}. Loads the promise + call graph, suggests whisper/transfer/hold, records orchestra intent with source_type=promise for agent handoff. Managed SaaS runtime. ≤120/hour/key (shares orchestra bucket). Idle $0. carrier/pstn always false. No cross-tenant graph merge. Advanced: not Asterisk-first glue; own PBX optional footnote.

Mark an open promise kept or broken with an optional privacy-safe note — makes the sales demo writable, not read-only. ≤60/hour/key → 429. POST /v1/pbx/promises/close {promise_id, status:kept|broken, note?}. Phones/emails in note are scrubbed. Scoped by x-temsor-pbx-key when set. Idle $0 — no LLM/PSTN.

Export open/broken promises as CSV or JSON — hotel / PBX-provider unclosed-promise sales list. GET /v1/pbx/promises/export?status=open|closed&format=csv|json. CSV/JSON include closed_at, close_note, path crumbs. Scoped by x-temsor-pbx-key when set. Idle $0.

Undo a kept/broken close — restore an open promise so the sales demo can reverse a mis-click. ≤60/hour/key → 429. POST /v1/pbx/promises/reopen {promise_id}. Only kept|broken → open; clears closed_at/close_note. Scoped by x-temsor-pbx-key when set. Idle $0.

Privacy-first erasure (default-on SaaS): clear derived index by id/phone; durable audit who/what/when + tenant scope. ≤30/hour/key → 429. Managed pointer only — not a legal deletion SLA. E.3 Privacy-first redaction for managed SaaS. mode=anonymize|delete. Lookup by target_id or phone. Clears Person/Promise/Fact/partial index + Temsor audio pointers in your tenant only (no cross-tenant). Durable audit. Raw media stays on your side. Not a legal deletion SLA. List: GET /v1/pbx/redact/audit.

List recent erasure actions — who/what/when + tenant scope. Managed SaaS derived-index audit; not a legal deletion SLA. GET /v1/pbx/redact/audit?limit=20. Durable audit from POST /v1/pbx/redact (E.3). Tenant-scoped. Derived index cleared; raw media not deleted by Temsor. Number→flow managed SaaS; own PBX optional. No DID sales.

Register a pilot PBX; returns pbx_id + API key once (hashed at rest). ≤5/hour/IP → 429. Idle $0. Control-plane register. API key shown once; stored hashed at rest. Send x-temsor-pbx-key to scope GET /v1/pbx/promises to this pbx_id. No speech recognition while idle.

Agent archive search over call promises and facts — voice-Google for call-ops, not a receptionist bot. Idle $0. Demo index + domain dictionary v0. Filters: status, date, ext, competitor, stuck_after (transfer/handoff hop). Returns call-graph path + transfer edges for orchestration. Idle $0. No live PSTN.

Control-plane status — promise/orchestra/search counts + idle $0. No DID sales. GET /v1/pbx/status. Scoped by x-temsor-pbx-key when set. Derived index on Temsor for agent search/orchestration. Idle $0. Advanced: own PBX/Asterisk stays on customer.

Register or list outbound webhooks — customer URL receives promise/fact/orchestra/redact events (HMAC if secret). POST write ≤30/hour/key → 429. POST /v1/pbx/webhooks {url, secret?, events?, ping?} registers. GET lists. ping:true sends a test delivery. Events: promise.opened, promise.closed, fact.extracted, orchestra.act, redact. Signed with x-temsor-signature when secret set. Idle $0. No DID sales. Erasure-aware. Advanced: own PBX/Asterisk stays on customer.

List last N outbound webhook deliveries (stub or live) — promise/fact/orchestra/redact/ping. GET /v1/pbx/webhooks/deliveries?limit=20&webhook_id=. Durable log of Temsor→customer POSTs. HMAC when secret set. Idle $0. No DID sales. Erasure-aware. Advanced: own PBX/Asterisk stays on customer.

Capability: mode=stub|ready from TEMSOR_RTP_*. No live RTP on forge; idle $0. GET /v1/pbx/whisper/status. Reports MediaBridge config presence (TEMSOR_RTP_PATH). mode=ready means path env set — not that Asterisk RTP is flowing. live_rtp/listening always false on forge (R2–R3 IVR sidecar; R4 lab E2E harness). No ≤400ms production claim. TEMSOR_RTP_LAB=1 → optional lab_latency_ms on suggest. No DID sales.

Live agent handoff cards (≤3): open promises/facts from in-call index. Idle $0. Lab: TEMSOR_RTP_LAB=1 → lab_latency_ms. POST /v1/pbx/whisper/suggest {call_id, partial_text}. Returns open promises and facts as whisper cards for agent handoff. Prefers in-call partial index when present. Mock latency_ms under 400ms (compat stub). Idle $0. No Telnyx/DID/live PSTN. Advanced: optional lab_latency_ms when TEMSOR_RTP_LAB=1 (measured; not prod SLA).

Validates an IBAN from any of 70+ countries: checksum, country length and in-country account structure, and resolves the bank and branch codes. Most free libraries stop at the mod-97 checksum. That only says the digits are self-consistent — not that the number could exist in that country. This checks the country's BBAN structure too, so a wrong-length or wrong-shaped account is rejected before your payment file reaches the bank. When only the two check digits are wrong, the correct ones are computed and returned as a suggestion, because that is what people actually mistype. Turkish, Dutch and Belgian IBANs additionally resolve to a bank name when the national code is in our table; otherwise bankName is null. This is not a SWIFT directory.

Builds a Turkish IBAN from a 5-digit bank code and account number, computing ISO 7064 check digits. A Turkish IBAN is TR + 2 check digits + 5-digit bank code + 1 reserved digit + 16-character account. This pads the account on the left, defaults the reserved digit to 0, computes the mod-97 check digits, and resolves the bank name from the same national table used by IBAN validation when the code is known. Unknown bank codes still produce a structurally valid IBAN with bankName null — existence at the bank is not claimed.

Builds a UBL-TR e-Invoice or e-Archive XML document from plain JSON, computing every total and validating the parties. Selling into Türkiye means issuing a UBL-TR document whose element order is fixed by schema and whose totals must agree to the kuruş, or the integrator rejects it. This endpoint takes the invoice as ordinary JSON and returns the XML. Totals you send are ignored on purpose — line amounts, per-rate VAT subtotals and the payable amount are all recomputed here, because a rounding difference of one kuruş is the most common rejection. VKN and TCKN checksums are verified, and the amount is written out in Turkish words as invoices require. It does not sign the document and does not transmit it: the financial seal and the submission to the tax authority belong to your certificate and your integrator. What comes back is a document ready to enter that step.

Turns a UBL-TR e-Invoice or e-Archive XML document into clean JSON: parties, line items, taxes and totals. Works regardless of the namespace prefix the sender used (cbc:, cac:, ns0:), normalises single-line documents into arrays, and reports amount mismatches in `warnings` instead of returning quietly wrong totals. The job that costs accounting and expense software the most engineering time.

Splits a free-form Turkish address into neighbourhood, street, building, floor, flat, district, province and postcode. Handles the abbreviation chaos (Mah./Mh., Cd./Cad., Sk./Sok., No:12/5, K:3 D:7), cross-checks the province against the postcode, repairs misspelled district names against a dictionary, and returns a confidence score. Anything it could not place is listed in `unparsed` — nothing is dropped silently. Built for shipping, checkout and CRM systems that receive Turkish addresses typed by humans.

Adds business days or counts them between two dates, accounting for Turkish public and religious holidays including half-day eves. Ramadan and Sacrifice feasts follow the Hijri calendar and cannot be derived reliably by formula, so announced dates are read from a table; years without an official announcement are returned with `confirmed:false` rather than guessed silently. For delivery promises, SLA clocks, payment terms and shipping estimates.

Checks an e-mail address: syntax, whether the domain can actually receive mail, disposable and role-account detection, and typo correction. No SMTP probing. Most competitors connect to the target server and try RCPT TO, which sounds thorough but produces a confident-looking number you cannot trust: most servers are catch-all, the big providers refuse to answer, and doing it in bulk gets the probing IP blacklisted. We report only what is provable — whether the domain has a mail exchanger, whether it is a known disposable service, whether the mailbox belongs to a role rather than a person, and whether the domain looks like a typo of a common one (gmial.com → gmail.com). Addresses are also normalised, so Gmail dots and +tags stop creating duplicate signups.

Computes Turkish severance (kıdem), notice (ihbar) and annual-leave entitlement from service dates and the gross wage, using the official ceiling and minimum-wage tables for the given day. Payroll and HR tools in Turkey chase a parameter that changes every January and July: the severance ceiling, the SGK cap, the minimum wage. This endpoint applies the statutory formulae (Labour Law 4857 arts. 17 and 53, former 1475 art. 14) to those tables. The ceiling is applied to the monthly wage, not the total. Stamp tax (0.759%) is deducted from severance; income tax is not — kıdem is exempt. Notice pay IS taxable; we return the gross and say so, because the actual withholding depends on the employee's cumulative tax base. What this will not tell you: whether the employee is entitled to severance at all. That depends on the reason for termination (retirement, just cause, marriage, military service…). Treating the number as "what is owed" is how you lose at trial.

Writes a monetary amount out in Turkish words, the way invoices, cheques and promissory notes require. Applies the rules that trip up generic libraries: Turkish says "bin", never "bir bin"; the kuruş part is read separately; and both "1.234,56" and "1,234.56" are accepted and told apart automatically. A mandatory field on Turkish e-invoices, cheques and notes — with no off-the-shelf API until now.

Builds a person-to-person TR Karekod (FAST) payment payload from IBAN, name and optional TRY amount. Text only — no PNG. Turkish instant payments (FAST) can start from a TR Karekod that encodes the payee IBAN and name in EMV-style TLV. This builds the person-to-person organisation (payload format indicator 750210): application template with IBAN, payee name, optional 12-digit kuruş amount, optional free-text explanation, and a CRC-16/CCITT-FALSE trailer. The Turkish IBAN is normalised and checked with ISO 7064 mod-97; non-TR IBANs and non-TRY currencies are rejected. This does not send money and does not render a QR image.

Validates and normalises any phone number to E.164, with structural checks for every country calling code. For signup and checkout flows that need to store one canonical form and reject typos early. Any common input (spaces, dashes, brackets, 00 or + prefix, a stray national trunk zero) comes back as E.164 plus national notation, with the country resolved from the calling code. Numbers without an international prefix use `defaultCountry` (US when omitted), so (415) 555-0134 and +1 415 555 0134 store the same. One thing this endpoint deliberately does not claim: the current mobile operator. Numbers are portable in most markets, so a prefix says where a number was first allocated, not where it lives today. Competing APIs report the prefix owner as "the operator" and customers pick SMS routes on that basis. Where we know the original allocation we return it as `originallyAllocatedTo` with the caveat attached, because a confident wrong answer costs more than an honest gap. Countries without a full numbering plan here are validated structurally and say so in `warnings`.

Validates a 5-digit Turkish postal code and resolves the province from the first two digits (plate map). PTT postal codes are five digits; the leading pair matches the province plate code for the 81 provinces. This endpoint checks the shape, maps 01–81 to the province name, and returns null with confidence none when the prefix is outside that range — it does not invent a district. District-level resolution needs the full PTT directory and is out of scope here.

Computes the deemed-received date and the HMK deadline (with holiday shifting) for a Turkish notification, from the date the underlying event actually happened. Turkish notification law (7201) ties the deadline clock to an event that is not "the date on the letter" — for electronic notification the UETS platform reports send, read AND reached dates, and only the reached date starts the clock (art. 7/a: deemed received 5 calendar days after reaching the address, whether or not it was opened). HMK adds two more rules on top: the day of notification itself does not count (art. 92 — the period starts the next day) and if the computed last day lands on a weekend or a full public/religious holiday, it moves to the next business day (art. 93); a half-day eve (arife) does not shift it. What this will not tell you: whether the notification was itself valid, or what a specific `periodType` (itiraz, temyiz, cevap…) is in days for your case — that number differs by statute and we do not guess it; supply `periodDays` yourself. Only the "hmk" law family is covered so far.

Validates Turkish national ID, tax number, IBAN, licence plate, IMEI, barcodes, KEP, MERSİS, postal code and ISBN from one endpoint, with type auto-detection. Goes past a yes/no: resolves the bank behind an IBAN, the province behind a licence plate and the GS1 country prefix behind a barcode. Pure local computation — no upstream service is called, so latency is microseconds and the answer never changes for the same input.

Splits a Turkish amount into net, VAT and (optionally) withholding: who pays the seller, who remits the withheld VAT. For invoice lines and checkout totals that have to show KDV dahil / hariç and, when the parties are in a withholding (tevkifat) situation, the split between what the buyer pays the seller and what the buyer remits to the tax office. Rates are 0, 1, 10 and 20 percent. Withholding is a fraction of the VAT (2/10 through 10/10), or a named code from the GİB partial-withholding list. The 2026 threshold (TRY 12,000 gross) is reported and compared with the amount, but `applies` is always null: whether withholding actually applies depends on the taxpayer status of both parties, which this endpoint does not know. Feeding it a fraction is not a legal opinion.