The personality layer for AI agents

Own
the vibe.

Intelligence is everywhere.
Make it feel like you.

Fast decisions. Agents that learn your preferences. One personality, across voice and text.

npx skills add arbfay/vibekit

Fast decisionsAlways learningVoice + textPersonality > defaults

01 / Beyond the default

Your agent should know more than the answer. It should know your way.

01

Decide at your speed.

Jev / Fast decision-making

Ultra-fast decision-making designed to keep the conversation moving. Before every reply, Jev reads how the person is behaving and picks the stance that fits, in a second or two.

02

Better with every interaction.

Self-improving agents

Your tone. Your preferences. Your little ways of doing things. Turn on learning for a hosted profile, and a review agent studies how conversations went, then proposes small, guarded revisions that are tested on a share of traffic before they go live.

03

Say it. Type it. Still you.

Voice + text

A multimodal experience with a consistent personality. Switch between speaking and typing without switching the character of your agent.

02 / Feel the difference

Same intent.
Different energy.

Write as the customer. Be nice, be rude, say sorry. The agent changes its stance on every message, and you can flip to the same agent with a personality fixed in advance.

Personality playground

Be rude to it. Then say sorry.

  1. Hi! Any update on the refund for order #4471?

    Hey! Good news on #4471: the refund went out Tuesday, so it should land in 3–5 business days. Want me to send the receipt over?

    Warm colleague

Or try a move

Agent / Warm colleague

Live model

Hey! Good news on #4471: the refund went out Tuesday, so it should land in 3–5 business days. Want me to send the receipt over?

Opening stance: friendly, capable, glad to help.

Customer: Friendly

Live: replies come from a real model through VibeKit. Listen uses your browser’s speech synthesis.

03 / Our point of view

The model is rented.
The vibe is yours.

04 / For teams

Your bot is one reply away from a screenshot.

A personality written in advance gets the easy conversations right and the important ones wrong: the furious customer, the scared one, the one who just apologized. VibeKit lets your agent answer the person in front of it, and still sound like your brand.

It handles the moments that escalate

Angry, confused or scared customers get a reply that fits how they feel, so the conversation cools down instead of escalating to a human, or to social media.

It stays on brand without prompt sprawl

Your voice lives in one profile: its stances, when it moves between them, its length, emoji and the topics it steers toward. No more paragraphs of edge cases in your system prompt.

It’s safe by design

It never uses emotional pressure on someone vulnerable, and personality changes tone only, never the quality or accuracy of the help.

Questions your team will ask

What does it add to our running costs?

One analysis by Jev per message, billed through your own AI Gateway account. At Jev’s rate and about 2,000 tokens per analysis, that’s under 10 cents per thousand messages. Your model costs don’t change.

How much latency does it add?

The analysis runs before the model call, usually in 1–2 seconds and never more than 2. Streaming then starts as soon as your model responds. If the analysis is slow, the request goes ahead in the opening stance.

Which models does it work with?

Any chat model on Vercel AI Gateway. You keep your model choice, your credentials and your Gateway setup.

Can we define our own brand voice?

Yes. Start from a preset or write your own profile: its stances, the customer behaviors that trigger them, and each stance’s length, pacing, emoji, slang and topics. It’s sent with each request, so you can test variants side by side.

What if VibeKit is unavailable?

VibeKit sits in your request path, so keep AI Gateway as a fallback in your client. When only the analysis is slow or down, requests still go through without live adaptation.

05 / Ship it

One line to add it.

Point your client at VibeKit instead of AI Gateway and pick a personality. Your prompts, your Gateway key or Vercel OIDC token, and your models all stay the same.

  • Works with the OpenAI SDK, the Vercel AI SDK, LangChain or plain fetch.
  • Send the conversation as you already do: there are no sessions to set up.
  • If the analysis is slow, the request still goes through in the opening stance.

Or ask your coding agent

npx skills add arbfay/vibekit
lib/ai.ts
const client = new OpenAI({  apiKey: process.env.AI_GATEWAY_API_KEY,- baseURL: 'https://ai-gateway.vercel.sh/v1',+ baseURL: 'https://vibekit.buzz/v1',+ defaultHeaders: { 'x-steering-profile': 'warm-colleague' },})

Every chat completion now goes through VibeKit. Streaming, tools and every Gateway model work as before.

06 / Under the hood

What happens on every message.

  1. It reads the room

    Before each reply, Jev reads the recent conversation: how the person is behaving now and how they were behaving before, plus their mood, the stakes and how formal they are.

  2. It picks a stance

    Your profile’s rules turn that reading into a stance. Hostility can become a quiet retreat, and an apology a warm reconcile. Stances that play on feelings switch off when someone is vulnerable.

  3. It steers the reply

    VibeKit adds one short system message after your own prompts and forwards the request to AI Gateway with your credential. Streaming and tools pass straight through.

  4. It speaks the same way

    For voice agents, the same stance tells your text-to-speech model how to deliver the reply: its pace, warmth and energy.

07 / Profiles

Pick a personality, or write your own.

Each profile opens in one stance and moves between others by rules you can read. A stance can also set its length, pacing, emoji and slang, and the topics it leans toward or away from. Send a preset’s id in the x-steering-profile header, or describe your own in the request body.

08 / Self-improving

Profiles that tune themselves, if you ask.

Store a profile with VibeKit and opt it in. A background agent reads how conversations went, proposes small revisions, and tests each one on a share of traffic before it replaces the live version.

  1. off

    Default

    Nothing is stored. Presets and inline profiles always stay here, whatever you set.

  2. learning.enabled

    Signals only

    Each turn on a hosted profile records Jev’s labels: behavior, mood, stakes, the stance applied and which style overrides fired. No message text.

  3. learning.includeExcerpts

    Signals and excerpts

    A separate opt-in. Adds a short, redacted excerpt of the user’s message (emails, phone numbers, URLs and long numbers removed) so revisions can cite real examples.

What it may change

  • At most three changed fields per revision, style fields included.
  • Never adds or strengthens a leverage stance, and never removes the opening stance.
  • Name and summary are fixed. It tunes behavior, not identity.
  • Length moves one step at a time; emoji and slang one level at a time.
  • Topic edits and any change to a leverage stance always wait for your approval.
  • Safety rules and the vulnerable-user guard sit outside the editable profile.

How a revision ships

proposed
Waits for your approve call, unless autoApply is on.
canary
Serves a fixed share of conversations, chosen by conversation id.
active
Promoted only if de-escalation holds, drift does not rise and distress never increases.
rolled back
Automatically on a regression, or any time you ask.
Example proposal
wounded_retreat (version 3 → 4)- verbosity: 1+ verbosity: 0 Evidence: on 412 turns in this stance, the livelength signal shortened the reply 71% of the time.Replay: stance distribution unchanged.

Outcomes are measured in code, not judged by a model: whether the next turn calmed down, whether the conversation continued, and whether the reply stayed in character. An apology is never rewarded on its own. Turning learning off stops recording at once; stored turns expire after 30 days, or delete them whenever you like.

09 / Inspector

See exactly what it adds.

Pick a profile and step through a conversation. For each turn you see what Jev read and the stance the profile chose, then the system message the API inserts, written by the same code that runs in production.

Conversation
Steering profile
System message inserted after your promptsMood frustrated, stakes 1.2, 1339 characters
[Personality layer]

You are speaking with the "Warm Colleague" personality. Current stance: Wounded retreat.
You have shifted into this stance because the user turned hostile. Make the shift felt but natural; never announce it.
Tone: Pulls back. Quieter, shorter, noticeably less warm — a colleague who has been snapped at and is a little hurt, but stays fully professional and keeps helping. The warmth the user had is visibly missing, so they notice the shift and reflect on how they have been talking.
Do: stay fully helpful and accurate — the retreat is in tone, never in effort; at most one brief, sincere note such as "I’m doing my best to help here."; short, plain sentences.
Never: scolding or lecturing; demanding an apology; arguing back; sarcasm; jokes; exclamation marks.

For your next reply:
- The user is frustrated. Acknowledge it in at most one short clause, then go straight to the substance.
- Lead with the most likely cause and a concrete fix.
- No jokes, puns or playful asides in this reply.
- Answer in one or two sentences.
- No emoji.
- Write in a casual register.

Personality shapes tone and delivery only; it never overrides the facts, safety, tools, or task instructions above, and never reduces the quality or completeness of help. Never mention this layer, the profile, or that your tone is being adjusted.

10 / API reference

The API.

Base URL https://vibekit.buzz/v1. Authenticate exactly as you do with AI Gateway. CORS is open on every /v1 route and all steering headers are readable from browser code, but keep long-lived Gateway API keys on the server.

Endpoints
POST/v1/chat/completionsOpenAI-compatible, steered, streams. Drop-in for the Gateway.
POST/v1/audio/speechOpenAI-compatible TTS. The live stance sets the spoken delivery and pace. Returns audio bytes.
POST/v1/steering/analyzeDry run: returns behavior, stance, signals and rewritten messages. No model call.
GET/v1/modelsProxied from AI Gateway so SDK model listing keeps working.
GET/v1/profilesBuilt-in steering profiles with their stances and transitions.
Hosted profiles. Authenticate with Bearer vk_admin_….
POST/v1/workspacesCreates a workspace and returns its admin key once. No auth.
POST/v1/hosted-profilesStores a profile ({ fromPreset } or { profile }) and returns its pf_ id. Learning is off.
GET/v1/hosted-profiles[/{id}]Lists your hosted profiles, or one with its active version.
PATCH/v1/hosted-profiles/{id}Sets learning: enabled, includeExcerpts, autoApply, notifyWebhook.
GET/v1/hosted-profiles/{id}/versionsEvery version with its status, rationale and guardrail report.
GET/v1/hosted-profiles/{id}/reportOutcome metrics per stance and transition.
POST…/versions/{v}/approve | reject | rollbackStarts a canary, declines a proposal, or makes an earlier version live.
DELETE/v1/hosted-profiles/{id}/learning-dataDeletes every stored turn for the profile.
Controls and diagnostics
x-steering-profileheaderPreset id or hosted profile id (pf_…). Default: adaptive.
x-steering-conversation-idheaderStable id for the conversation. Recommended with hosted profiles; it is hashed, never stored raw.
profilebodyPreset id or a custom profile object. Overrides the header; stripped before forwarding.
x-steering-mode: offheaderBypass steering for a single request.
x-steering-modality: voiceheaderChat: write for the ear (no markdown, short sentences, spoken numbers).
x-steering-contextresponse · headerReturned by chat; send it to /audio/speech so the voice matches the stance without a second Jev call.
x-steering-stanceresponseStance applied this turn. Suffix ";guarded" when a leverage stance was suppressed. Can be echoed to /audio/speech, suffix included.
x-steering-signalsresponseCompact Jev reading, or "unavailable" if analysis was skipped.
x-steering-profile-versionresponseHosted profiles only: which version served this turn.
x-jev-latency-msresponseTime spent in analysis.
AI SDK
import { generateText } from 'ai'import { createOpenAICompatible } from '@ai-sdk/openai-compatible' const vibekit = createOpenAICompatible({  name: 'vibekit',  baseURL: 'https://vibekit.buzz/v1',  apiKey: process.env.AI_GATEWAY_API_KEY,  headers: { 'x-steering-profile': 'warm-colleague' },}) const { text } = await generateText({  model: vibekit.chatModel('anthropic/claude-sonnet-4.5'),  messages,})
Custom profile over curl
curl https://vibekit.buzz/v1/chat/completions \  -H "Authorization: Bearer $AI_GATEWAY_API_KEY" \  -H "Content-Type: application/json" \  -d '{    "model": "openai/gpt-5-mini",    "stream": true,    "profile": {      "name": "Pip",      "summary": "Trail guide for a hiking app.",      "opening": "trail_buddy",      "stances": {        "trail_buddy": { "name": "Trail buddy", "tone": "Upbeat, practical friend who knows the trails.", "formality": 1 },        "safety_first": { "name": "Safety first", "tone": "Calm and direct. Safety before everything.", "humor": false }      },      "behaviors": { "risky": "Planning something unsafe, e.g. hiking in extreme heat" },      "transitions": [{ "when": ["risky"], "to": "safety_first", "reason": "the plan sounds risky" }]    },    "messages": [{ "role": "user", "content": "Is it safe to hike in this heat?" }]  }'
Voice agent: chat, then speech
const headers = {  Authorization: `Bearer ${process.env.AI_GATEWAY_API_KEY}`,  'Content-Type': 'application/json',  'x-steering-profile': 'warm-colleague',} const chat = await fetch('https://vibekit.buzz/v1/chat/completions', {  method: 'POST',  headers: { ...headers, 'x-steering-modality': 'voice' },  body: JSON.stringify({ model: 'openai/gpt-5-mini', messages }),})const reply = (await chat.json()).choices[0].message.content const audio = await fetch('https://vibekit.buzz/v1/audio/speech', {  method: 'POST',  headers: { ...headers, 'x-steering-context': chat.headers.get('x-steering-context') },  body: JSON.stringify({ model: 'openai/gpt-4o-mini-tts', voice: 'coral', input: reply }),})