What’s new

Smart-MCP uses Jev, Kilevel's System One model, to decide which capabilities an AI is shown, request by request, instead of showing it everything at once.

• For your flows. Turn on Smart-MCP in a flow's MCP Servers section and mark each tool, Skill and MCP server as Always or Smart. Smart items are offered only on the turns where they are relevant, and the assistant can ask for anything that was held back. Try a message shows what a message would be offered, with relevance scores. See Smart-MCP.
• For Kilevel's own MCP server. A new endpoint, /api/mcp/smart, shows an AI client only the Kilevel tools that fit what it is connected for. Pick a preset (build, operate, analyze, data) or describe the work in a sentence, and the client finds the rest with discover_tools. See Smart-MCP.

A training is now a training module, and a module can have several personas: each one a caller of its own, written as a new role-play scenario, taken from one of your flows, or an AI agent.

• Who calls next. Pick how a persona is chosen for each practice call: at random by weight, one per person (an A/B test across people), in turn, in turn after a success (easy to hard), least practised first, or weakest first.
• Compare them. Every persona shows its calls and how many met the rule, over every task that assigns the module. A persona can be switched off without losing its results.
• Due when they joined. A training task can be due a number of days after each person is assigned, so people added later get the same time as the first group.
• Duplicate. Modules and tasks can be copied from their edit window or the table, as a new form to adjust before creating.
• Editing a task now shows the teams and people already on it, apart from the ones you add.

A new metric pack, Script Adherence, measures how closely agents followed the call script they were given — sixty measurements across required points, mandatory and prohibited statements, steps and order, branches and decision logic, questions, deviations and recovery, and how naturally the script was used.

• Measured against your own scripts. Choose the call script on a transcription job, or on a human call campaign — one for every stage, or a different one per stage. A campaign call is held to the script for the stage it was made in.
• A round of its own. The script metrics run in a separate analysis round with the script in the context, after your other outputs, and never without a script.
• Fair by design. A context-weighted score separates what the agent did from where the customer took the call, and metrics that did not apply are left empty rather than guessed.
• A Script Adherence scorecard preset weighs adherence in context, mandatory statements delivered and prohibited ones avoided.

See Analysis and metric packs.

The transcription guide became five pages — sources and watching, analysis and metric packs, details and linking and the API — and Performance has a guide of its own.

The API reference now carries a full description for every transcription endpoint: what the request body holds, why externalId makes a retry safe, and why the details you sent and the analysis a model produced are never merged into one object.

A new page for how every agent is doing — across calls, chats and imported recordings together, because an agent's work is not split by where it happened.

• Ten views over one set of filters: the agents table, a daily briefing written as prose, an overview, a roster for one-to-ones, a scorecard matrix, a dense workbench, team comparison, calibration, alert rules and privacy settings. Each view can be rearranged, and an arrangement saved — optionally with its filters, which makes it a saved view you come back to.
• Scorecards turn metrics into one weighted score and a pass mark, with per-criterion targets and auto-fail criteria reported as their own rate rather than averaged away. Presets are offered for common shapes, and an account with no scorecard is given one built from the metrics it actually produces.
• Review & calibration measures whether the scores hold up: draw a random sample per agent, score a handful by hand, and read the agreement, the typical gap, any bias, and the criteria the two sides disagree about most.
• Alert rules are questions asked on your behalf every night — a metric, a scorecard score, an auto-fail rate or conversation volume, per agent or account-wide. The editor says before you save whether a threshold can never fire, always fires, or would fire for half the floor.
• Teams, so people are compared against their own team's benchmark rather than only the account average.
• Privacy and retention: names, stable pseudonyms or aggregate-only figures, minimum group sizes, logged reveals for cause, and a retention limit on per-person history — built so the page can be run the way a works council agreement permits.
• Briefing by email, daily or weekly at an hour you choose, and Export PDF for whatever is on screen.
• New automation triggers: New Conversation Score, New Transcription Analysis, Performance Alert and Performance Alert Resolved.

The page and the metric packs are fully translated, German included. See Performance.

• S3-compatible buckets — Amazon S3, Cloudflare R2, MinIO, Wasabi, Backblaze B2, DigitalOcean Spaces and Google Cloud Storage. The cheapest source there is: each recording is handed to the transcription provider as a short-lived signed link to your own bucket, so the audio never passes through us. Read-only keys are enough.
• FTP, FTPS and SFTP — a folder on your own file server, read at the moment each recording is transcribed, with an optional host key fingerprint and sub-folder support.
• Watching that survives a growing folder. A watched source can move or delete each recording once it has been transcribed, skip files that are still uploading, pause importing above a queue you choose, and warn you when nothing has arrived for longer than you expect — the one failure a watch cannot otherwise report.
• Which speaker is the agent is worked out per recording: the channel, then a self-introduction, then the model, then a fallback you choose — and which of them answered is recorded. A job can also carry a default call direction.
• Redaction removes emails, phone numbers, IBANs, card numbers and the names you imported from the transcript before it is stored or analysed, with a preview before you commit.

See Sources and watching.

A metric pack is a ready-made structured output, written and locked so a metric means the same thing in January and in September and on every agent. Twelve ship with the platform: a core conversation-quality pack, sector packs for sales, support, collections, healthcare, insurance, financial advice, technical support, booking and retention, and two coaching reads.

• Attach them anywhere an output goes — transcription jobs and flows alike. They sit in the same list as the outputs you write yourself.
• Written reasons. Judgement fields carry a short reason citing what was actually said. The model works the reason out before committing to the answer, so the answers come out better, not merely more explainable.
• Measured metrics need no analysis at all: talk share, dead air, interruptions, speaking rate and longest monologue, taken from the utterance timings.
• Classifications — resolution status, churn risk, sentiment, outcome — are reported as rates rather than dropped for having no average.
• Structured outputs gained icons, colours and a declared direction, so a chart knows which way is good without being told twice, and judgement scales widened from 1-5 to 1-10.

See Analysis and metric packs.

Imported audio cannot say who the agent was or which customer called, so a job can now be pointed at where you already keep that.

• Details from the file name or a spreadsheet — map a segment of the name, or a column of your CSV, to the agent, the customer, the conversation id and time, project, stage, goal and tags, or to fields you define yourself. The setup screen shows what would be attached, worked through a real example, before anything is transcribed.
• Linking ties a recording to an agent or a contact in your workspace — agents on an exact id or name, contacts on the normalised phone number only. Optionally creating the ones that are missing.
• A transcription API: configure a job once, then push recordings into it by URL or as an upload, with the details the calling system already holds. externalId makes retries free.

See Details and linking and Pushing recordings over the API.

• Artifacts preview in place — a Work's deliverables open inside the run rather than needing a download, and their content is streamed authenticated rather than through a public link.
• Dead sandboxes recover themselves. A sandbox that has gone away is rebuilt and the rebuild is reported to the model, so a long Work stops failing silently halfway through.
• Identical deliverables are stored once, matched on content, so a run that writes the same file twice no longer produces two.
• Test a Work-mode agent from the agent manager, the same way you quick-test a calling agent.

Simulated calls and chats — the ones you place yourself while building — are now visible only to whoever ran them, everywhere: call lists, contact centre views, job counts, status summaries and the API. They no longer move anybody else's numbers.

Runs from before the change were backfilled, so a colleague's test calls disappear from your lists rather than only new ones.

• Jump nodes reach a distant node without dragging a wire across the canvas. A Jump behaves exactly as a wire drawn straight to its target; it never runs and has no output of its own. Available in both the automation editor and the state graph, and existing connections can be converted in bulk.
• Connector styles per canvas: curved, rounded, right angle, metro, straight or arc — with an option to keep parallel lines apart on the orthogonal ones.
• Evaluations are drawn inside their state card, so a graph with judgement on every state stays readable.
• State gating now covers skills, knowledge, structured outputs and MCP servers, not just tools — so what a state may reach is decided in one place.

Kilevel now has public documentation, with guides for building, launching and monitoring agents, integration and administration guides, and an API reference generated from the live service.

Every endpoint in the reference has a built-in playground: add your API key once and send test requests straight from the page.

• Browser calls can be monitored live, the same way as phone calls.
• Speech is formatted per language before it is spoken with ElevenLabs voices, so numbers, dates and abbreviations are read the way a native speaker would say them.

• Bulk export: download every transcript of a job as a ZIP, or combined into a single file.
• Watched Drive folders: point a transcription job at a Google Drive folder, and new audio is transcribed as it arrives.
• Transcript layout: view transcripts two-sided or stacked, and open them full screen.
• Duplicate detection across jobs is now opt-in.

Transcribe audio recorded outside Kilevel — uploads, Google Drive files or URLs — and run the same structured outputs a call gets, with results in the same shape as a call record.

Speakers are separated by channel or by voice, and multilingual recordings in which speakers switch languages are transcribed in one pass. See Transcription.

• Collect User Input is a new automation step: send a person a form, with a stable public link, and continue the automation with their answers. Forms can be generated by AI from what the step needs to know.
• Agent tasks can now carry context items, be scheduled, and return structured results.

• MCP servers can be picked from a catalog and connected with stored credentials, and used in flows as well as automations. Servers with many tools are loaded progressively, and their resources can be pinned into context. See Connecting MCP servers.
• Process Video is a new automation step for combining and editing video, and image and video generation steps accept reference images.
• Generated media is kept as a durable, downloadable artifact.

• Start an AI conversation from an automation: place an outbound call or chat with a flow or agent, and monitor it live while it runs.
• Benchmarks compare models and settings across the same scenarios, with a cost estimate before each run.
• Automations and benchmarks are now billed per run, with a monthly free quota.

The usage page has a new Overview tab showing every limit on your workspace — calls, chats, automation runs, agent tasks — how much of each is in use, and what raises it. Capacity for automations, agent tasks and website chats now scales with your concurrency lines. See Concurrency and limits.

Fish Audio is now available as a voice provider — an EU/GDPR-oriented option for workspaces that need one. See Voice and language.

• Barge-in controls set how readily the assistant stops speaking when the caller starts.
• Speaking when unaddressed can be set per state, with custom instructions for what to say.
• Acknowledgements during tool waits: the assistant tells the caller it is working on something instead of going silent, and can say a short preamble before calling a tool.

• AI disclosure: an opening sentence telling callers they are speaking with an AI, as required by Article 50 of the EU AI Act. Always on for EU workspaces, available per flow elsewhere, with preset wording in every supported language or your own. See AI disclosure.
• Recording consent: new flow nodes to ask for consent, start and stop recording — recording starts automatically on a yes. See Recording consent.