Skip to main content
Glama

Uhura

Your AI assistant can now make the phone call.

AI assistants can research, write and plan, but some things still only happen on the phone: the doctor's office without online booking, the travel agency that answers emails slowly, the shop that won't say on its website whether it's open on Friday.

Uhura closes that gap. Ask Claude, or any assistant that speaks MCP, to make the call: it drafts the briefing, you approve it, and an AI voice agent places the call. The agent introduces itself as an AI and works through your questions. If it lacks an answer, it asks you while the person waits on the line. You get the transcript, and what the call cost.

Uhura is a small self-hosted service in front of ElevenLabs Agents. It works from your AI assistant through MCP, or from the command line. One instance can serve one person on a laptop or several people from a server.

Status (2026-10-04): built and verified against a real ElevenLabs account. Real phone calls work through a Twilio number, in German, including a company's phone menu and a waiting queue the agent sits through silently; progress reports and the cost of each call are verified on real calls too. Asking you a question during a spoken call has so far only been verified in text conversations, and no English call has been made yet. See docs/verified.md.

How a call works

  1. Draft. Give the number, the language, the goal and the facts the agent may share. Uhura checks the number and shows the exact briefing. Nothing is dialled.

  2. Rehearse (optional). Talk to the real agent in text, playing the person being called, to see how it handles the briefing.

  3. Confirm. A separate step dials.

  4. Follow and steer. The agent reports progress as it goes (menu, queue, person reached, answers). If it lacks an answer, it asks you and waits on the line for your reply. Instructions you send reach it with its next check-in, at the latest just before it hangs up.

  5. Result. The transcript, duration and cost in ElevenLabs credits are stored; no audio is kept.

$ uhura draft --to "030 23125 000" --principal "Erika Musterfrau" --language de \
    --topic "eine Frage zu Ihren Öffnungszeiten" \
    --goal "Ask whether the shop is open on Friday."
Message composed, Captain. Awaiting your order to transmit.
  call id : 74b8167429cf
  number  : +493023125000
  opening : Guten Tag, hier spricht ein KI-Assistent im Auftrag von Erika Musterfrau. …
  after menu: Guten Tag, ich bin ein KI-Assistent … Es geht um eine Frage zu Ihren Öffnungszeiten. …

$ uhura rehearse 74b8167429cf      # text conversation, nobody is called
$ uhura confirm 74b8167429cf       # dials, follows the call, prompts for answers
Hailing frequencies open.

The brief can be in English; the agent speaks the call's language (--language, German here) and always opens with the fixed disclosure line for it. The topic completes the sentence "Es geht um …" in the line the agent says to the first person it reaches after a phone menu or queue, so it is written in the call's language.

Related MCP server: Dialfetch MCP

Responsible use

Uhura places real phone calls to real people, with an AI voice. It is meant for calls you would otherwise make yourself: asking a shop, a doctor's office or a travel agency something on your own behalf. It is not meant for marketing, mass calling, surveys, or anything that hides that an AI is calling or on whose behalf.

The guardrails below are built for that use: every call opens with a fixed disclosure, each call is drafted and confirmed separately, and the limits keep volume low. They are not a licence to call anyone about anything. You are responsible for the calls you place and for the laws that apply to them. docs/legal-notes.md records the reasoning for Germany only; it is not legal advice, and other countries' rules on automated calls, recording and transcription differ.

Guardrails

Enforced by the service, whatever the briefing says:

  • Drafting never dials, and a draft can be confirmed only once.

  • Emergency, premium-rate, shared-cost and service numbers are refused.

  • Only countries in UHURA_ALLOWED_COUNTRIES can be called.

  • Per person: UHURA_DAILY_CALLS calls per 24 hours and UHURA_MONTHLY_MINUTES per 30 days. A call is only dialled if the minutes left cover the longest call the agent will make (20 minutes); calls still running count at that length until their result arrives.

  • Per person: UHURA_DAILY_REHEARSALS rehearsals per 24 hours, each closed after 10 minutes.

  • The opening line is fixed per language. It says that an AI is calling, for whom, and that the call is transcribed, and asks for agreement. A briefing cannot change it.

  • People only see their own calls. Calls and transcripts are deleted after UHURA_RETENTION_DAYS.

  • When the agent reports that the person objected, no transcript of the call is kept, neither by Uhura nor at ElevenLabs; only the refusal, its time, duration and cost.

  • The service does not start with placeholder or short tokens or tool secret (fewer than 16 characters), since it is usually reachable from the internet.

Asked of the agent through its fixed rules (reliable in tests, but a language model follows them, it is not forced to): information only, no bookings or payments; report it and end the call if the person objects to transcription; ask instead of inventing facts; in phone menus, press keys toward the goal and never agree to a recording; stay silent on hold; give the fixed, short re-introduction (shown in the draft) to the first person reached after a menu or queue; keep turns short and do not repeat the briefing.

Quick start

uv sync
uv run uhura init           # creates .env with a fresh token and tool secret
                            # then add your ElevenLabs key to .env
uv run uhura setup-agent    # creates the agent; put the printed id into .env
uv run uhura serve          # http://127.0.0.1:8787
uv run uhura check          # tells you what is still missing

The full walk-through, including the phone number, the public address the agent's tools need, Docker and hosting for several people, is in docs/setup.md.

Commands

Command

What it does

uhura init [--name NAME]

Create .env with a fresh user token and tool secret (never shown, never overwritten)

uhura serve

Run the service

uhura setup-agent

Create or update the agent in your ElevenLabs account

uhura setup-tools [--url URL]

Point the agent's tools at the service's public address (default: local ngrok tunnel)

uhura check [--rehearsal]

Verify the setup; with --rehearsal, run a scripted conversation

uhura draft …

Prepare a call and show the briefing

uhura rehearse ID

Try a draft in text against the real agent

uhura confirm ID

Dial, follow the call, answer its questions

uhura watch ID

Follow a call that is already running

uhura say ID "text"

Queue a follow-up instruction

uhura show ID / uhura list

Transcript, status and cost / recent calls

Use from an MCP client

claude mcp add --transport http uhura http://localhost:8787/mcp \
  --header "Authorization: Bearer <token>"

The service serves the tools itself at /mcp. For clients that can only start a local program there is a stdio adapter, uhura-mcp; see docs/operations.md.

Tools: draft_call, confirm_call, wait_for_event, answer, send_instruction, get_call, list_calls, rehearse_call, rehearse_say, end_rehearsal. The server tells the client to show every draft to the user and to confirm only after approval.

Documentation

Development

uv run pytest          # no network needed
uv run ruff check .    # lint
uv run ruff format .   # format

The name

Lieutenant Nyota Uhura is the communications officer of the USS Enterprise in Star Trek, played by Nichelle Nichols. She opens channels, hails other ships and relays what they say to the captain. Uhura does the same job for you: it places the call, passes the agent's questions to you and your answers back. Its status lines, from "Hailing frequencies open." to "Channel closed.", are in her voice and live in src/uhura/phrases.py. The ribbed earpiece in the banner is a nod to the one she wore on the bridge.

License

MIT.

Available Tools

10 tools
answerC

Answer a question the voice agent asked mid-call. question_id is the event's seq.

Inputs: text (required), call_id (required), question_id (required). Returns: object

confirm_callA

Dial a drafted call. Only use after the user approved this specific draft.

Inputs: call_id (required). Returns: object

draft_callA

Prepare a call without dialling. Returns the call id, the normalised number, the fixed opening line, the voicemail message and the full briefing the voice agent will receive. voicemail is what to say if a mailbox answers; a fixed AI introduction is spoken before it. Without it the agent hangs up on voicemail. topic is a short phrase completing "Es geht um …" / "It is about …" (e.g. "eine Reiseanfrage"); the agent says it in its fixed re-introduction to the first person it reaches after a phone menu or queue. No full stops or question marks. progress switches the agent's progress reports ('progress' events) on or off for this call; by default the service decides.

Inputs: to (required), goal (required), principal (required), facts, +6 more. Returns: object

end_rehearsalC

End a rehearsal. The draft stays unchanged and can still be confirmed.

Inputs: call_id (required). Returns: object

get_callB

Status, duration, transcript and cost of a call. cost (once the call has ended) has the ElevenLabs credits charged, the billed minutes after the silence discount, and ElevenLabs' dollar prices for the voice platform and the language model.

Inputs: call_id (required). Returns: object

list_callsC

The user's recent calls, newest first.

Inputs: limit. Returns: object

rehearse_callA

Start a rehearsal of a drafted call: the real agent with the real briefing, in text, with nobody being called. Read what the agent says with wait_for_event, passing the returned events_after as after ('agent_said' events), and reply as the person being called with rehearse_say. Questions and instructions work as in a real call.

Inputs: call_id (required). Returns: object

rehearse_sayC

Say something to the agent in a running rehearsal, as the person being called.

Inputs: text (required), call_id (required). Returns: object

send_instructionA

Queue a follow-up instruction for a running call. It reaches the agent with its next tool call, at the latest just before it hangs up, so delivery is not immediate.

Inputs: text (required), call_id (required). Returns: object

wait_for_eventA

Wait up to timeout seconds for new events on a call (questions from the agent, 'progress' reports with stage and note, call ended). Pass the highest seq seen so far as after. Returns an empty list if nothing happened; call again until status is 'done' or 'failed'.

Inputs: call_id (required), after, timeout. Returns: object

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    B3.2/5.0

    Scored across 10 tools

    Disambiguation4/5

    Most tools have clearly distinct roles: draft_call prepares, confirm_call dials, rehearse_call/rehearse_say/end_rehearsal form a coherent rehearsal group, and get_call/list_calls are reads. Minor overlap exists between answer (real-call question), send_instruction (real-call follow-up), and rehearse_say (rehearsal reply), but descriptions clarify the boundaries.

    Naming Consistency4/5

    Most names follow a predictable verb_noun pattern (draft_call, confirm_call, get_call, list_calls, send_instruction, rehearse_call, rehearse_say, end_rehearsal). The bare verb 'answer' and the noun-less 'wait_for_event' are minor deviations from the dominant convention.

    Tool Count5/5

    Ten tools is well-scoped for a voice-call agent covering drafting, dialing, live interaction, and rehearsal. Each tool earns its place with no obvious redundancy.

    Completeness4/5

    The lifecycle is well covered: draft, confirm, event-waiting, answering, instructing, and inspecting calls, plus a full rehearsal loop. Gaps are minor, such as no way to edit or cancel a draft, but core workflows are complete.

    Maintenance

    ActivitySlowing
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables AI agents to make real-world phone calls with AI voice technology and provides tools to track call status, transcripts, and summaries. It supports automated communication with both live numbers and simulated businesses for testing and demonstration purposes.
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables AI agents to place real phone calls, navigate IVR trees, and retrieve structured answers with transcripts and recordings.
      MIT
    • A
      license
      A
      quality
      C
      maintenance
      Enables an AI assistant to place user-approved phone calls in the callee's language through your own Bland or Twilio account, with language detection, script generation, cost estimates, and safety guard rails. Includes a preview-then-confirm gate, plus call status, transcripts, and summaries.
      8
      MIT