User Research API: Embed AI Interviews into Any Product or Workflow

The Bottom Line

A User Research API lets you trigger AI-moderated user interviews from your own backend code instead of a SaaS dashboard. Koji exposes a complete headless API: start an interview, exchange messages, complete it, and read the analyzed transcript back. There are no outbound webhooks yet, so your backend polls the read endpoint or listens for the embed widget's browser events. Either way you can embed conversational research anywhere a customer touchpoint exists: an in-product onboarding flow, a churn cancellation screen, a post-purchase email, a Discord bot, or an internal data pipeline.

The core primitives are four REST endpoints plus a JavaScript embed widget that posts completion events to the parent page. With these you can run a full customer research conversation programmatically: the AI interviewer adapts to each participant, asks follow-up questions, and your backend reads quality-scored transcripts and structured answers as soon as analysis finishes.

This guide is the entry point. If you're evaluating whether a research platform fits your stack — or you've outgrown survey APIs that only POST form responses — this article shows how Koji's API differs and where to go next.

Why a User Research API Exists

Most research tools assume a researcher logs into a dashboard, sends a link, and reads the results manually. That model breaks at three common moments:

1. In-product moments matter more than scheduled studies. The best feedback comes from a user mid-task, not a user three weeks later trying to remember what frustrated them. A research API lets you trigger an interview the moment a user hits a meaningful trigger (clicks "Cancel subscription", finishes onboarding, downgrades, etc.).
2. Product-led growth teams want every signup interviewed. When you're onboarding hundreds of users a week, sending them all the same Typeform isn't research — it's a survey. An API lets you queue an AI interview per qualified signup and feed the structured output into your CRM.
3. AI-native teams want research data as a feed. You want themes, quotes, and structured answers landing in your data warehouse, Slack, or your own LLM agent — not a PDF.

A real user research API answers all three. Koji's does. Closed-dashboard tools like SurveyMonkey or Qualtrics cannot — their APIs let you POST survey responses, not run conversational interviews.

The Core API Primitives

1. Headless Interview Endpoints

The Koji REST API exposes four conversation endpoints under /api/v1:

• POST /api/v1/interviews/start — Open an interview session. Accepts the projectSlug (or projectId), optional respondentMetadata (name, email, external ID), and returns a sessionToken plus the first AI message.
• POST /api/v1/interviews/message — Send a participant message and receive the next AI response. The AI agent decides whether to ask a follow-up, move to the next question, or wrap up.
• POST /api/v1/interviews/complete — Mark the interview complete and trigger analysis. The transcript is quality-scored and themes are extracted asynchronously.
• GET /api/v1/interviews/{id} — Read the full transcript, quality score, structured answers, and themes after analysis runs.

Every endpoint is rate-limited at 60 requests per minute per API key, with permissive CORS for browser-based callers. Authentication is via Bearer tokens — keys are generated in the API tab of any Koji workspace.

2. The Embed Widget (JavaScript)

If you'd rather not implement the full conversation loop yourself, Koji ships a JavaScript embed widget that renders the interview UI inside your app. Drop the snippet into a modal, a sidebar, or a dedicated page; it handles voice, text, and the structured-question widgets (buttons, sliders, radio, checkbox, drag-to-rank) automatically. The widget is the same primitive that powers Koji's hosted interview page — your branding, your domain, with the AI moderation logic still running on Koji's backend.

3. Completion Signals for Near-Real-Time Pipelines

Koji has no outbound webhooks yet, so there is no URL to register and nothing signed arriving at your server. Two patterns give you the same outcome:

• Poll the read endpoint. Store each interview_id when you start an interview, then have a scheduled job call GET /api/v1/interviews/{id} and act when status is completed and analysis is present, which is when the quality score, themes and structured answers are all populated.
• Forward the embed event. If the interview runs in the embed widget, the respondent's browser posts koji:interview_completed to the parent page. Forward it to your backend as a fast path, and keep the poller as the source of truth: the event only fires while the tab is open.

This is how teams build "research-aware" automations: post every interview transcript to a Slack channel; sync structured NPS answers into HubSpot; trigger a Linear ticket when a churn-risk theme appears; feed insights directly into a Claude agent that drafts PRDs. The trigger is your own job, not a Koji callback. See Webhook Setup for working code.

4. CRM and CSV Import

The respondent import endpoint accepts CSV uploads or programmatic POSTs. Each respondent gets a personalized interview link (with a unique token) so you can target named accounts, send 1:1 invites, and tie interview results back to a contact record in your CRM. When you need people you do not have, the same study can also be filled from Koji's built-in panel: open Panel recruit in the dashboard, describe the audience, approve the price per completed interview in credits, and those interviews land in the same study, readable through the same GET /api/v1/interviews/{id} endpoint as everything else.

Authentication and Rate Limits

Koji uses standard Bearer-token authentication:

Authorization: Bearer koji_live_<your-api-key>

API keys are scoped to a workspace. Each key has a creation date, last-used timestamp, and revocation control. There are no per-endpoint scopes today — every key has full read/write access to its workspace.

Rate limits:

• 60 requests/minute per API key (sliding window).
• No daily cap. You're bounded by credits, not requests.
• CORS is permissive on /api/v1/* endpoints — direct browser calls work, useful for client-side embed scenarios.
• No outbound webhooks yet. There is no signed callback to validate. Poll GET /api/v1/interviews/{id} or forward the embed widget's koji:interview_completed event instead.

If you exceed the rate limit, you'll get a 429 Too Many Requests with a Retry-After header.

Pricing and Credits via the API

The API and MCP connector are available on every plan, including a free account, and outbound webhooks are not available yet. Pricing is credit-based and identical to dashboard interviews:

• Text interview: 1 credit
• Voice interview: 3 credits
• Report refresh: 5 credits
• Quality gate: interviews scoring 1 or 2 (out of 5) don't consume credits

Interviews start as low as €1 per qualified interview, and €3 per qualified voice interview. The API and MCP connector are included for everyone, not sold as a tier, and there are no per-seat or per-call API surcharges. When you need more credits, top up with pay as you go. Nothing is ever charged automatically, and you pay only for the interviews your study actually uses.

For team-wide API deployments (multiple products embedding interviews, SSO/SAML, isolated infrastructure, and a signed DPA), the Enterprise plan adds the controls procurement asks for, with custom committed credit volume and invoicing.

Compared with survey APIs that charge per-response or per-question, Koji is unusually simple: a research-grade text interview costs as low as €1, and a conversation that scores below 3 out of 5 is free and never charged.

Architecture Patterns

Here's how typical implementations look.

Pattern A: In-Product Cancel Flow

A SaaS product wants to interview every user who clicks "Cancel my subscription." The cancel button posts the user to a modal that loads the Koji embed widget with a pre-filled respondentMetadata (user ID, plan, tenure). The AI runs a 3-minute interview asking why they're leaving, what almost stopped them, and what would bring them back. The transcript is analyzed; a poller on your side reads it back and, if the theme "missing feature" appears, posts the quote to the product team's Slack.

Pattern B: PLG Onboarding Research at Scale

A developer-tools company wants to interview every signup that completes the tutorial. A backend cron job runs daily, identifies qualified signups, hits POST /api/v1/interviews/start to generate personalized links, and sends them via email. A second scheduled job polls the read endpoint and routes finished transcripts into the company's warehouse (BigQuery) tagged with the user's plan and traffic source. The product manager queries the table for emerging themes by cohort.

Pattern C: Headless Research for an AI Agent

A founder builds an AI agent in Claude that supervises customer research. The agent calls Koji over MCP (Model Context Protocol) to read existing studies, draft new ones, generate reports, and import respondents. From the founder's perspective, customer research is a tool the AI uses — not a SaaS the founder logs into. Koji's MCP server (15 tools) is built on top of the same API; everything callable via REST is also callable via MCP.

Pattern D: Embedding Research in a Mobile App

A mobile team wants in-app feedback that goes deeper than NPS. They use the embed widget inside a sheet view, configured for voice interviews. Users speak into their phone; the AI converses, probes, and wraps up in 90 seconds. The team sees themes ranked by frequency, quality-scored, with verbatim quotes — all without anyone scheduling a call.

How Koji's User Research API Compares

The TL;DR: survey APIs collect answers; recording APIs save videos; Koji's API runs conversations.

Getting Started

1. Create a workspace. Sign up at koji.so. You get 10 free credits immediately.
2. Generate an API key. Settings → API → Create Key. Copy it once — it's only displayed at creation time.
3. Build your first study in the dashboard. Easier to validate the brief in the UI before going headless. Set interview mode (structured, exploratory, or hybrid) and add structured questions if you want chartable answers.
4. Test POST /api/v1/interviews/start. Use cURL with your key and the study's projectSlug. You'll get a sessionToken and the first AI greeting.
5. Loop messages. Send a participant reply via POST /api/v1/interviews/message and read the AI's next message. The agent may attach a widget field (for scale, single_choice, etc.) so your client can render the right input.
6. Complete and read results. Call POST /api/v1/interviews/complete to finalize. Analysis runs asynchronously (typically under 30 seconds). Poll GET /api/v1/interviews/{id} for the quality score and structured answers, or forward the embed widget's completion event if the interview runs in the widget.

For production, run the poller on a schedule and put the embed widget on a dedicated subdomain. The full API specification, including request/response schemas, is in Headless API Overview.

Frequently Asked Questions

Can I use the API on the free plan? Yes. The API and MCP connector are available on every plan, including a free account, and outbound webhooks are not available yet. Voice and text interviews run on every way of paying, and new accounts get 10 free credits to try interviews in the dashboard.

How do I authenticate? Bearer tokens. Authorization: Bearer koji_live_<key>. Keys are workspace-scoped and revocable.

What's the rate limit? 60 requests/minute per key, sliding window. Returns 429 with Retry-After on overflow. Most production workloads stay well below this.

Are voice interviews supported via the API? Yes. Voice interviews cost 3 credits each. You can either use the embed widget (handles audio capture for you) or implement the full WebSocket-based voice protocol yourself.

How do I know when an interview has finished? There are no outbound webhooks yet, so there is nothing to register or verify. Poll GET /api/v1/interviews/{id} from a scheduled job and act when status is completed and analysis is present. If the interview runs in the embed widget, forward the browser's koji:interview_completed event to your backend as a fast path. See Webhook Setup.

Can I use Koji headlessly with no dashboard at all? Yes. Every dashboard action — create study, edit brief, publish, generate report, export data — is available via the REST API and the MCP server. Several customers operate fully headless via Claude + MCP.

Related Resources

• Headless API Overview — full REST specification and code samples
• Starting Interviews via API — step-by-step POST /interviews/start walkthrough
• Completing Interviews via API — finalization, analysis triggers, and result retrieval
• Embed Widget Reference — JavaScript widget configuration
• Research Automation: building near-real-time research pipelines
• Webhook Setup: why there are no outbound webhooks yet, and the three patterns that replace them
• API Authentication — Bearer token issuance and rotation
• Structured Questions Guide — the 6 question types returned in API responses
• MCP Integration Overview — same primitives, exposed to AI agents

Further reading on the blog

• B2B Customer Research: The Complete Guide for Product Teams (2026) — B2B customer research is harder than B2C — you are navigating buying groups of 10+ stakeholders, gatekeepers, and enterprise procurement cyc
• Customer Research Done Right: A Complete Guide for Product Teams — Customer research is the foundation of every successful product decision. Learn the types, methods, and best practices that help product tea
• Customer Research for Product-Led Growth: The Complete Guide (2026) — Product analytics tells you what users do. It cannot tell you why they activate, churn, or expand. This guide covers the research questions,

<!-- further-reading:blog -->