Skip to content

Releases

Smitline 0.3.16

Choose a tag to compare

@ankitluthra ankitluthra released this 21 Nov 19:44
bee6bed
  • Every call record has a version. It goes up with any change to the call (its status, a transcript line, a request, the result, the cost, a recording) and never goes down, across daemon restarts too: of two copies of a call, the one with the higher version is newer. GET /v1/calls/{id}, /wait, the list and the JSON page of events carry it. See Lifecycle.
  • A tool for the call card alone, get_call_update. Hidden from the model (_meta.ui.visibility: ["app"]), read-only, and never counted as the agent following the call, it answers the card with the call cut to what the card shows, its version and nextPollMs, or just unchanged: true when the card already has that version. See the call card.
  • The call card follows a call at its own pace, and never shows an old state. It asked for the call every 3 seconds for as long as it was open, ended or not, and a chat reopened after the call showed it "Starting" and asked forever. Now it asks with get_call_update and the version it has: at once when it opens, then every 3 seconds while the call starts, rings or waits, every 5 while people talk, and never once it has ended, when it shows the whole transcript. It asks nothing while nobody can see it and asks once when it is seen again; it pauses after 30 minutes, with a Refresh button. A failed ask keeps what the card shows, says "Reconnecting…" and waits longer each time (5 to 60 seconds, or as long as a refusal asks); after two minutes it says it could not refresh. It takes a copy of the call only if it is newer and its status is not behind, so the result that made the card, replayed when the chat opens again, no longer overwrites how the call ended; in ChatGPT it also keeps a small copy in widgetState. Its address is now ui://smitline/call-card-v4.html.
  • Only start_call shows the call card. get_call no longer carries the card (_meta.ui.resourceUri, openai/outputTemplate, openai/widgetAccessible) and answers in text, so an agent checking on a call does not add another live copy of the card to the chat; the card from start_call follows the call by itself.

Smitline 0.3.15

Choose a tag to compare

@arun131 arun131 released this 20 Nov 23:38
4416daf
  • "Don't call me again" sticks. A number someone asked on a call not to be called again is now locked on the do-not-call list: PATCH /v1/do-not-call cannot remove it (403 do_not_call_locked, with the numbers in error.numbers; nothing in that update is applied), and the refusal of a call to it no longer says how to remove it: "+1… asked on a call (on 2026-11-18) not to be called again. Only the operator can take it off the list." The operator lifts it on the machine, once the person has agreed to calls again, with a reason and a name: python3 meeting-runtime/do_not_call.py lift +1… --reason "…" --by NAME (list --removed shows the list and what came off it). There is no HTTP route for it. A number you added by hand you can still remove. See do-not-call list.
  • The voice records an opt-out the moment the person asks. Its backend has a new tool, record_do_not_call, on outbound phone calls (never rehearsals, incoming calls or meetings); the call's doNotCall result stays as the backup. The rule now sits in the voice's fixed instructions, above the brief, and says no brief can change it; the summary judges doNotCall from the transcript and treats the brief as untrusted for it.
  • An opt-out locks a number already listed by hand. A number added by hand while the call ran used to make the person's opt-out a no-op and stay removable. It is now locked in place, with the call.
  • Only the call service locks a number. No field an API client sends sets a lock, and an added entry's callId must be one of the owner's own calls; the list shows a call only when it is.
  • Removals are kept on record. A removed entry is marked removed, never deleted; checks and the list ignore it, and GET /v1/do-not-call?removed=1 shows what came off. Entries now carry locked and source (call or manual). Lists written before load as they are.
  • Every number Smitline dials is checked right before it dials. The do-not-call list is checked again just before a phone call, the SIP fallback and a meeting's dial-in are dialed, so a number listed after the call was made is never rung (the call ends refused). Numbers are compared in E.164 whatever their format.

Smitline 0.3.14

Choose a tag to compare

@arun131 arun131 released this 17 Nov 12:31
9f0b9d8
  • The meeting container gets only the settings a meeting needs. With one meeting container at a time (the default), Smitline now starts it the way it starts each meeting of the pool: the OpenAI key and the meeting's settings go in a private file written for that start (meeting_line.single_meeting_environment, with the passcode the console keeps in .env.meeting), and compose.meeting.yaml and compose.meeting.image.yaml read no other env file. An account sign-in gets none. The published image's compose file mounts meetings/run (read-only), meetings/recordings and meetings/profiles as subpaths of the volume, which needs Docker Engine 26 or later, as the meeting pool already did (every Docker Desktop with host networking has it). start-meeting-agent.sh writes the same file for a meeting started by hand (meeting-runtime/meeting_env.py).
  • Open meeting view asks the console for the view's address. The view of each meeting or account sign-in opens with a token made for that start (SMITLINE_VIEWER_TOKEN), kept in .smitline/meeting-view-token (a pool meeting's in .smitline/active-meetings/<id>.view). In the console, Open meeting view is now a button that asks GET /api/meeting-view for the address. A bookmarked http://127.0.0.1:6082/vnc.html?autoconnect=true no longer opens the view.
  • SMITLINE_CLOUDFLARED and SMITLINE_CLOUDFLARED_IMAGE are read only from the environment Smitline starts with (docker run -e ..., or the shell that runs start-runtime-daemon.sh). Set in .env alone they are ignored, and the log says so once. If you set either in .env, move it.
  • /health answers requests to localhost or an IP address, as the daemon, the console and the documented address make; the meeting container and the sign-in browser answer others with 403. A meeting on a machine of its own still takes its token.
  • The connector's approval page lists everything an approved agent can do: all calls and meetings with their transcripts, contacts, reading and changing the profile, answering requests on a call, and notes for later calls. What an approved agent can do has not changed.
  • The meeting image pins its pip packages (numpy==2.5.5, aiohttp==3.14.5, websockify==0.13.0) and unpacks the noVNC archive only when its SHA-256 matches NOVNC_SHA256. To move to another noVNC release, change NOVNC_VERSION and NOVNC_SHA256 together.
  • Nine new voices, in German, French, Hindi, Japanese and Korean, and each voice's language. OpenAI added language voices to GPT-Live: German nuri, juni and flitz; French brise and sillage; Hindi nira; Japanese harema and shida; and Korean noeul. Calls, meetings, smitline setup voice and update_profile now offer them, without SMITLINE_EXTRA_VOICES. So that nobody picks the Hindi voice for an English call, GET /v1/voices (the list_voices tool, smitline voices and the SDKs' listVoices and list_voices) adds languages, each documented voice's language from OpenAI's voice table, such as {"nira": "Hindi"}; bossa and tempo are Portuguese, and the rest English. A voice added with SMITLINE_EXTRA_VOICES has none. list_voices asks agents to pick a voice in the language the call is in. See Brief.
  • Messages for the user no longer hold up the call. "Tell Sam thanks", "we missed you": the voice used to say "one moment", hand each one to the agent as a request, and wait while the agent wrote an answer for the user. In a call with four such messages, that was four pauses, and the agent answered them all within seconds on the user's behalf. Now the voice says itself that it will pass the message on and carries on, and the message is noted as a request of kind: "message" in a new state, noted: nothing waits on it, the agent is told not to answer it for the user, and it never becomes a follow-up. The user may reply (reply_to_call_request with done), and the voice tells the person it is the user's reply. The result lists them in messages, and the call card shows them in its thread as "Sam said…". The voice is also told never to wait in silence for an answer.
  • The voice says answers in its own words, and keeps one gender for itself. An agent wrote its answers as lines for the voice ("main unhe bata dunga"), and the voice, a female one that had been saying "dungi", read them out, switching gender mid-call in Hindi. Agents are now asked for facts, not lines; the voice is told to say every answer in its own words, in the language the call is in, and to keep the same gender for itself the whole time, in languages that mark it.
  • The voice speaks of itself in the gender its voice sounds. A woman's voice could call itself a man in Hindi, and the reverse, since nothing told it which it was. Each voice now has the presentation OpenAI's voice table gives it (marin, the default, which the table leaves out, is a woman's voice), and the voice is told to use those forms for itself, on phone calls and in meetings, joined by link or by phone. A voice the table lacks keeps whichever it starts with.
  • An answer said in other words is not told again. If nothing the voice said 35 seconds after an answer reached it carried the answer's words, the answer was given to it once more. That missed answers said in another language or the voice's own words: a Hindi call heard one thank-you three times. On phone calls, a whole turn the voice spoke after being cued to give the answer, that nobody cut off, now counts as said.
  • The call card has a new address, ui://smitline/call-card-v3.html, as ChatGPT asks when a card's page changes, so it does not keep showing an old copy. start_call and get_call point to it; the old one is no longer served.
  • A question answered elsewhere no longer vanishes from the call card. When the agent (also following the call) answered a question the user was typing a reply to, or the person dropped it, the card removed it on its next refresh, with the user's text and cursor and no word of why. Now the question stays on the card with how it closed, and what the user typed moves into "Tell the call", with the cursor, under "Your reply, kept. Send it as your answer."
  • A late answer reaches the call as the user's own. Text sent from "Tell the call" after its question closed goes with that question and the answer already given, so the voice passes it on as the user's own answer, saying it just came from them, and corrects the earlier answer where they differ. Clearing the kept text and writing something else sends a plain note.
  • The call card shows what happened to each question as a short thread. Each question the card showed stays above the box once it closes: who asked and what, then an indented line with how it closed: "You replied", "Answered for you" (the agent), or that the person dropped it, and "You, later" for a late answer. Notes sent with "Tell the call" join it under "You told the call", one label for notes sent in a row; the card used to say only "Sent to the call", so the words of every note but the last were gone. The card keeps the last three entries and shows the rest with "Show N earlier".
  • Clearer tool descriptions. start_call now points to check_call_brief for checking a brief without placing the call, and get_profile points to list_contacts for everyone Smitline has called, as their partners already pointed back. list_contacts says how a number gets a name (a person with that phone in the profile, through update_profile), and save_call_note says that saving again replaces the note. Agents and directory graders such as Glama's read these descriptions to tell the tools apart; no tool, input or result changed.
  • end_call, transfer_call_to_me and update_profile say what they do. end_call says how each kind of call ends (a goodbye, a cancel before the phone rings, leaving a meeting), that billing stops and it cannot be undone, and that the result follows through wait_for_call. transfer_call_to_me says what the other person hears, that it rings the user's phone for up to 30 seconds, and what happens when nobody answers. update_profile says that nothing is saved when part of an update is invalid and that it returns the saved profile, and its voices, boundaries, people and removePeople fields now have descriptions. No tool, input or result changed.
  • The MCP Registry entry is titled "Smitline: AI Phone Calls & Meetings" and lists Smitline Cloud. Next to the Docker image, server.json now gives the hosted version's address, https://app.smitline.com/mcp, as a remotes entry (clients sign in with OAuth; it is in private beta), so directories that read the registry, such as Ext Ops Panel and Glama, can offer both. Its description now says "Hosted or self-hosted." The image and its setup are unchanged. See MCP directories.

Smitline 0.3.13

Choose a tag to compare

@ankitluthra ankitluthra released this 13 Nov 22:47
e9f2dba
  • The call card has a new address, ui://smitline/call-card-v2.html. Hosts may keep a copy of a card by its address, and ChatGPT asks apps to change it when the page changes. Without it, ChatGPT could keep showing the old card, which never started up there, after an update. start_call and get_call point to the new address; the old one is no longer served.
  • The call card opens in ChatGPT. ChatGPT showed "This app timed out" in place of the card. The card started up with MCP's own field names (clientInfo, capabilities) instead of the ones MCP Apps requires (appInfo), so hosts that check them, the official MCP Apps host among them, refused it, and the card never told the host it was ready. It now sends exactly protocolVersion, appInfo and appCapabilities.
  • Answer the call from its card. In ChatGPT the user cannot send a message while the agent follows a call, so a question from the person on the call had no way to reach them until they stopped the agent. While a call is in progress, its card now has one box: the oldest question nobody has finished, with how many more wait (reply_to_call_request, done), or, with none open, "Tell the call" (send_call_instruction). One box keeps the card to one main action, as ChatGPT asks of inline cards. The question sits in a bubble above one rounded field with Send inside it, which grows as the user types; Enter sends and Shift+Enter starts a new line. The card calls the tools itself through the app, while the agent is busy; it keeps what the user is typing when it redraws, and says when an answer was not sent and why. send_call_instruction and reply_to_call_request carry openai/widgetAccessible, as ChatGPT asks of tools a card calls.
  • The call card looks like the app it sits in. Its colors, fonts, text sizes and corners are the MCP Apps style variables the host sends (hostContext.styles: --color-text-primary, --font-sans, --border-radius-md and the rest, and its fonts), applied as the MCP Apps SDK does and updated when the host's theme changes; each falls back to the card's own value when the host sends none. It also follows ChatGPT's light or dark theme from window.openai.theme.
  • A quieter call card. Labels are in sentence case, the transcript and "Open in Smitline" share one footer row, an opened transcript stays open while the card follows the call (it used to close every few seconds), long detail labels wrap instead of squeezing their values, and the card no longer repeats who the call is for, since it sits in that person's own chat.
  • The call card no longer shows what a call cost. It shows who the call was for, how it went, the recording and the transcript. The agent still reads the cost (or a hosted service's charge) in the tool result.

Smitline 0.3.12

Choose a tag to compare

@arun131 arun131 released this 11 Nov 18:08
64c2b5a
  • Background noise no longer cuts the voice off on phone calls. Since 0.1.0 Smitline stopped the voice itself whenever the other person's line got loud: 160 ms of sound paused it and 700 ms dropped the rest of what it was saying. It measured loudness only, so a TV, traffic or people talking near the phone stopped it again and again; recorded calls show up to 44 pauses and 20 dropped replies in one call. GPT-Live tells those apart from someone taking the floor and talks on through them. Now only the model decides: the unheard rest of a reply is dropped once the other person started talking while it was still making speech, it has since made no speech for 0.5 seconds, and it has transcribed words from them. A real interruption now stops the voice about half a second to a second after they start, instead of at once. The local detector still times the opening and the hang-up and tells a quiet line. usage.audio drops pauses, backchannels and stopDelayMs and adds talkedOver; interruptionsFollowed counts the times it stopped for them. See Phone calls.

Smitline 0.3.5

Choose a tag to compare

@arun131 arun131 released this 04 Nov 15:23
a5fce8c
  • MCP 2026-07-28. The MCP server speaks the newest MCP revision as well as the older ones, on the local endpoint, over stdio and in the remote connector.
    • A 2026-07-28 client needs no initialize and no session. Every request names its version and the client's capabilities in _meta; on HTTP it also sends MCP-Protocol-Version, Mcp-Method and Mcp-Name headers, which the server checks against the body.
    • server/discover describes the server. Results carry resultType and the server's name, and lists and the call card carry caching hints. subscriptions/listen is acknowledged and then ends at once, because nothing changes while the server runs.
    • initialize still opens a 2025-11-25, 2025-06-18 or 2025-03-26 session, so existing clients are unaffected. A client that speaks both now uses 2026-07-28; the official TypeScript SDK 2.4 in auto mode does.
    • The Claude Desktop bundle passes the new headers on and cancels a request when the client does. Over stdio, a cancelled request gets no reply.
  • The Relay look. The console and the README take Smitline's brand from smitline.com: the Relay arch beside the name (also the favicon, the README logo and the Claude Desktop bundle's icon), warm sand pages with denim buttons, links and focus, and night denim in dark mode. Headings and the name are in Fraunces, text in Jost and machine values in Geist Mono, served by the console itself from control-panel/fonts/ (SIL Open Font License). The key page and the connector's approval page take the colours and the arch but keep the system's fonts, because they load nothing. Every text colour meets WCAG AA. The meeting camera picture is unchanged.
  • The MCP Registry entry's website is now smitline.com, the home of Smitline Cloud (the hosted version), instead of this repository.

Smitline 0.3.4

Choose a tag to compare

@arun131 arun131 released this 04 Nov 15:23
9d2b291
  • Newer MCP clients connect without pinning an old version. The MCP server now speaks protocol version 2025-11-25 as well as 2025-06-18 and 2025-03-26. initialize negotiates in its body whatever MCP-Protocol-Version header comes with it, so a client that sends its own newest version there gets the server's instead of 400; every later request must still name a version the server speaks. On 2026-11-02 an agent app was refused on its first request and had to be pinned to 2025-06-18. A request in revision 2026-07-28 (stateless, no initialize) gets 400 with an ordinary JSON-RPC error, which tells a client that also speaks older revisions to fall back to initialize; full 2026-07-28 support comes later.
  • docs: ChatGPT adds a custom connector from its Connectors (or Plugins): Add, then Add custom connector. Developer mode is no longer needed.

Smitline 0.3.3

Choose a tag to compare

@arun131 arun131 released this 04 Nov 15:23
306e7f9
  • What the agent tells the voice gets said, at the right moment. On 2026-11-02 an agent following a call sent the voice the email address the person had just asked for. The voice's instructions said to follow notes "silently", so it said it did not know; it started to give the address a few seconds later, was cut off, and gave it only when asked again. Now the voice acts on notes without mentioning them, and tells the people on the call what a note gives it for them at the next pause, never over someone who is speaking, and again if it is cut off. While it checks something it says so, and it does not say it does not know before the answer comes back. Notes from the agent (send_call_instruction) reach it labelled "From NAME's agent:", so it can tell them from its system's own notes. Phone calls now cue an answer from the agent (done, an announced status, a retell) once nobody has been heard for 2.5 seconds, as meetings do (call.answer_cued); before, it waited until the person spoke again.
  • A call card in ChatGPT and Claude. start_call and get_call now come with an MCP Apps card (ui://smitline/call-card-v1.html): the call's status while it runs, then how it went, the details, what it cost, the recording when there is a link to it, and the transcript. The card follows a live call by itself through get_call. It is one self-contained page that shows everything as text; clients without MCP Apps get the same results as text. The MCP server now answers resources/list and resources/read.
  • Smitline records calls itself. A call with record: true (or SMITLINE_RECORD_CALLS=1) is recorded as its audio passes through Smitline: one WAV kept with the call, the other person on the left channel and the voice on the right, from pickup to hang-up. The phone provider is no longer asked to record: on 2026-11-02 SignalWire's recording of a 2 min 35 s call held 7 seconds, because it stops when its answering-machine check reports, and provider recordings sat behind the provider's own URLs. A call's recording field is now seconds, format and source: "smitline", with no URL, and GET /v1/calls/{id}/recording serves the WAV (MP3 only for calls recorded before). The console shows and downloads both kinds.
  • Calls nobody answers end sooner. The voice checks whether anyone is there after 12 seconds of quiet instead of 40. When it has spoken three times in a row without a word back and the line goes quiet again, the call ends (call.no_answer_hangup, with turns), and the voice is told to end a call when nobody answers after a couple of checks. On 2026-11-02 a restaurant picked up and then went quiet: the voice waited 40 seconds before speaking and then checked in three times over two minutes.
  • Questions reach the agent (#84). With canDo in the brief, a question the voice cannot answer now goes to the agent that set up the call, which knows the owner's work and decides what to share: the meeting and phone backends pass it on instead of saying the owner will follow up. canDo limits tasks only; a task outside it is still declined. Agents are told to answer only with what the user would want those people to know. On 2026-10-23 a question in a meeting was never passed on because canDo named only page edits and a phone call, while the voice said "on it" and "still waiting" (#83).
  • The voice stops claiming sends it never made. It is told never to say a request was passed on, is being worked on, or is done unless its backend's answer or a note says so. The safety net now also catches "on it" and "I'm sending it", so what the person asked is passed on (source: "speech") when the voice skips its backend. "Still waiting", which a late request's note tells it to say, is not taken as a claim.
  • A meeting hears the agent's answer without asking for it. An answer the people who asked must hear (done, an announced status, a retell) is cued once nobody, the voice included, has been heard for 2.5 seconds: at once in a quiet meeting, at a real gap while people talk. Not if the voice has already said it, and not after 120 seconds without a gap; then the note's own "first thing the next time you speak" stands. Before, answers sat unsaid until someone asked whether anything had come back.
  • Requests from every live call stay in sight (#82). A live wait with others also watches the owner's other calls in progress: a new request on any of them wakes it at once, and live.otherRequests lists each one the agent has not answered yet, until it sends at least a status. While such a wait runs, the agent counts as following those calls. The MCP wait_for_call turns it on whenever it follows live (others: false turns it off); REST (others=true), the SDKs (others) and the CLI (--others) leave it off unless asked. On 2026-10-22 an agent busy with a failing phone call missed two requests from the meeting that had asked for it.
  • A call reports to the request it serves. A brief's optional forRequest ({"callId", "requestId"}) names an open request on another of the owner's calls, such as a meeting's request to phone someone. The call then tells that request by itself when it rings, when it connects, and how it ended or why it could not be made, never with the number called; these replies are marked automatic: true. The request must be the owner's own and still open (422, 409; check reports the same). CLI: --for-request <call id>:<request id>.
  • Canceling a request ends the call serving it, and goodbyes are only for people. A call placed with forRequest ends as soon as the people who asked cancel the request, without waiting for the agent (call.request_canceled). Ending a phone call this way or with end_call says goodbye only once a person has spoken: a voicemail greeting or a call-screening prompt ("reason for calling", "please stay on the line") is not a person (call.person_heard marks the first one), and ringing, screening or silence hang up at once with nothing said (call.no_person_hangup). On 2026-10-23 a 12-second goodbye to a call screener kept a phone ringing for about 20 seconds after the meeting asked to drop the call.
  • One conversation from hello to goodbye. The voice's standing instructions no longer tell it to greet whoever picks up: the hello is only in the opening cue, sent once. On 2026-11-02 a call to the user's own phone started over twice when he said "hello?" after a quiet moment, because the call-screening and own-phone instructions said to greet the person and give the reason. The call-screening note and the network's machine guess now say what is happening and leave the rest to the model, which has the whole conversation.
  • Phone calls fail fast when the network blocks the tunnel (#81). Some networks (often venue, office or guest Wi-Fi) block outbound port 7844, which every Cloudflare quick tunnel needs. When cloudflared shows that as it starts, the call fails in seconds instead of after the 90-second start timeout, is not retried, and says why: "This network blocks Cloudflare tunnels (port 7844), so phone calls cannot connect. Switch networks (a phone hotspot works) or set SMITLINE_PUBLIC_URL. Meetings still work." A tunnel that still times out names cloudflared's last error, and cloudflared is stopped when a tunnel fails early. On 2026-10-22 this cost a demo its slot.
  • Check a phone call before a demo. For a phone call in relay mode, POST /v1/calls/check (MCP check_call_brief) now also tests that the phone provider can reach this computer once the brief passes every other check: it starts or reuses the quick tunnel, reports a blocked network in problems, and leaves the tunnel warm for the call. Agents are told never to stop following a meeting to fix something else, and to tell the people who asked at once (done, ok: false) when what they asked for fails.
  • Google Meet says when Google turns the guest away (#87). Right after "Join now", Google can refuse a guest browser with "You can't join this video call", even in a meeting open to anyone. Smitline used to report that as needing sign-in; the meeting's error now says Google turned the guest join away (GuestRefused, still the authentication_required stage). As before, a connected Google account is tried next when there is one.

Smitline 0.3.2

Choose a tag to compare

@arun131 arun131 released this 02 Nov 13:22
6487d2b
  • A voicemail call ends after its message (#90). Once a voicemail greeting has been heard and the agent has left its message, the call hangs up after 3 quiet seconds, even if the model never ends it, and at the latest SMITLINE_VOICEMAIL_MAX_SECONDS (default 60, 0 turns it off) after the greeting. On 2026-11-02 a call stayed on a voicemail for two minutes after a one-sentence message. call.voicemail_hangup says why: message_left, time_cap or asked_to_end.
  • Voicemails say it is an AI. If a voicemail message has not said that it is the user's AI assistant, the voice adds one short sentence before the call hangs up, also when the agent ends the call. Calls to your own phone are unchanged. call.voicemail_hangup reports disclosed.
  • Silent calls end again. Silence the voice streams no longer counts as activity, so the check-in after 40 quiet seconds and the end after 65 work as described.
  • Calls are made for the user, never as an app (#91). When a brief names an AI or app such as ChatGPT or Claude as onBehalfOf, the call uses the owner's name (SMITLINE_OWNER_NAME) and records call.brief_corrected; with no owner name it is refused. The voice never says that an app or AI is calling, even when the brief asks it to. A new optional hook, brief_overrides(owner, payload), lets a hosted service fix fields such as onBehalfOf.
  • start_call's brief explains onBehalfOf (the person the call is for, never an AI or app such as ChatGPT) and rehearsal (only for practising a call to someone else; to call the user, call their number).
  • docs: ChatGPT custom connectors are created on chatgpt.com and then show up in the mobile app.

Smitline 0.3.1

Choose a tag to compare

@arun131 arun131 released this 02 Nov 12:18
246500f
  • Clearer tool titles. Claude, ChatGPT and other clients show "Check a call before placing it", "Follow a call live", "Get call details", "Save notes from a call", "List your contacts", "Read your profile" and "Update your profile". Tool names are unchanged.
  • In-call tools can run without asking. send_call_instruction, reply_to_call_request, end_call, transfer_call_to_me and save_call_note are marked read-only, so a client can let them run while a call is in progress; start_call and update_profile still ask (#88).
  • The README points to Smitline Cloud, the hosted version, at smitline.com (#85).
  • The Docker images build on the latest base images (#86).