{"site":{"name":"Koji","description":"AI-native customer research platform that helps teams conduct, analyze, and synthesize customer interviews at scale.","url":"https://www.koji.so","contentTypes":["blog","documentation"],"lastUpdated":"2026-10-06T07:14:45.546Z"},"content":[{"type":"documentation","id":"189726b1-7599-4b0c-accd-42ac5224c6a9","slug":"chat-with-interview-transcripts-ai","title":"Asking Questions of Your Interview Transcripts with Koji and MCP","url":"https://www.koji.so/docs/chat-with-interview-transcripts-ai","summary":"Koji has no chat box over transcripts. The same job is done two ways. Inside Koji, every interview is analysed on its own and the report opens with the finding: key takeaways, themes with evidence counts, a section per question, personas and question coverage. Every claim carries a citation that opens an evidence panel with the supporting quotes and a link into the full transcript, so nothing has to be taken on trust. Outside Koji, the MCP connector at https://www.koji.so/api/mcp/mcp (OAuth, works with Claude, Claude Code, Cursor and any Model Context Protocol client) exposes koji_list_studies, koji_get_study_data, koji_get_interviews and koji_get_transcript, so an assistant can answer free-form questions, compare studies, surface disagreement, or draft a summary from the real conversations. Cross study questions are the main reason to connect it, since a report belongs to one study. Answer quality is set before the interviews run: capture the attributes you want to compare through imported personalised links or a screener, pair open questions with structured questions so numbers come from typed data, and let the quality gate keep interviews scoring below 3 out of the report.","content":"## The Bottom Line\n\nYou want to type a plain question, *\"Which participants pushed back on the proposed pricing?\"*, and get an answer with the verbatim quotes that prove it. That is a reasonable thing to want by the time a discovery practice has a few hundred transcripts in it.\n\nKoji does not put a chat box over your transcripts. It does two other things that between them cover the job.\n\nInside Koji, the answer is usually already written down. Every interview is analysed on its own, and the report leads with the finding, with every claim linked to the quote behind it. You click a claim, the evidence panel opens, you read the quotes, you jump into the transcript.\n\nOutside Koji, the [MCP connector](/docs/mcp-overview) hands your study data and your transcripts to Claude, Cursor or any MCP assistant. There you can ask anything you like in your own words, including questions that cross studies, and the assistant answers from the conversations it actually read.\n\nThis guide covers what each one is good for, how to connect the second one, the question patterns that get the most out of it, and how to set a study up so the answers are worth having.\n\n## What Actually Happens When You Ask a Question\n\nTake the pricing question above. Here is the honest path to the answer.\n\n**If the study is already analysed**, open its report. The themes section will have a pricing theme if one emerged, with an evidence count telling you how many participants it appears in. Click the claim and the evidence panel shows the quotes, with their participants, and a link into each full transcript. If the study asked a scale or choice question about price, the question section already has the distribution.\n\n**If you want it phrased as an answer**, ask an assistant connected over MCP: \"Read my Q3 pricing study and tell me which participants pushed back on the proposed price, with their quotes.\" The assistant pulls the interviews and transcripts through the connector and writes the answer.\n\nBoth routes end in the same place: real quotes from real conversations that you can open and read in context. Neither one is a search index, and neither one invents a participant.\n\n## Inside Koji: The Report Answers First\n\nIt is worth knowing exactly how much is already done for you, because most questions never need to leave the app.\n\n**Per interview analysis.** Each finished interview is analysed on its own: a quality score from 1 to 5, structured answers for each question, the themes in that conversation with their supporting quotes, a summary and a sentiment. Nothing is tagged by hand.\n\n**The report opens with the answer.** Key takeaways, then themes with evidence counts, then a section per question. See [Reading Your Research Report](/docs/reading-your-research-report) for what each section holds.\n\n**Every claim links to its quote.** Citations in the report open an evidence panel with the supporting quotes and a link into the transcript. This is the mechanism that makes the report checkable rather than something to believe.\n\n**Personas and question coverage.** The report groups participants into personas and shows how well each question in your brief was actually covered, so you can see where the evidence is thin.\n\n**The responses grid.** One row per respondent, one column per question, sortable by date or quality score. Fastest way to read every answer to one question.\n\n**Readable transcripts.** Every interview can be read end to end, with its structured answers and the interviewer's follow-ups in place.\n\n## Outside Koji: The MCP Connector\n\nFor free-form questions, and for anything that spans more than one study, connect Koji to your assistant.\n\n### Connect it\n\nThe connector lives at `https://www.koji.so/api/mcp/mcp` and authenticates over OAuth. Add it in Claude, [Claude Code](/docs/mcp-setup-claude-code), [Cursor](/docs/mcp-setup-cursor) or any Model Context Protocol client, sign in once, and the assistant can read your research. Setup per client is in the [MCP overview](/docs/mcp-overview).\n\n### The four tools that do the reading\n\n| Tool | What the assistant gets |\n|------|------------------------|\n| `koji_list_studies` | Every study you have, with status and interview counts |\n| `koji_get_study_data` | Aggregated study data with per-question statistics |\n| `koji_get_interviews` | The interview list with summaries, themes and sentiment |\n| `koji_get_transcript` | One full transcript with structured answers and analysis |\n\nThe assistant decides which to call. You just ask.\n\n## Four Question Patterns Worth Learning\n\n### 1. The disconfirming evidence question\n\n*\"Read my onboarding study and tell me who disagreed with the majority.\"*\n\n*\"Find the participants who did not want a dark mode and quote them.\"*\n\nThe highest leverage pattern, because synthesis naturally over-indexes on agreement. Ask for the dissent explicitly and you get the two-of-fifteen minority voice back. Inside Koji the same thing is visible in the low evidence count themes, which still carry their quotes.\n\n### 2. The cross study question\n\n*\"Across my churn studies from the last year, what is the most common reason people downgrade?\"*\n\n*\"Compare what people said about integrations in the enterprise study and the SMB study.\"*\n\nThis is the one thing the report genuinely cannot do, because a report belongs to a study. An assistant with the connector can list your studies, pull several, and reason across them. This is where connecting it pays for itself.\n\n### 3. The segment question\n\n*\"What did the design leads say about the new layout, compared with the individual contributors?\"*\n\nWhether this works depends on what you captured. Attributes you imported with personalised links, or collected in a screener or lead form, travel with the participant and show up in the analysis. Decide what you want to compare before the interviews run. See [Research Screener Questions](/docs/research-screener-questions).\n\nInside Koji, the personas section of the report answers the same class of question without you typing anything.\n\n### 4. The \"have we already asked this?\" question\n\n*\"Have I ever run a study that touched AI summarisation?\"*\n\n*\"What do I already know about onboarding for power users?\"*\n\nThis is how research stops being write-only. Ask before kicking off a new study. Half the time a previous one already answered the question, and the other half you start with grounded priors instead of from scratch.\n\n## Why the Answers Are Grounded\n\nThe fear with any assistant near research data is a plausible quote that nobody said. Three things keep this honest, and none of them are magic.\n\n**The assistant reads real transcripts.** `koji_get_transcript` returns the actual conversation. The assistant is summarising text it just retrieved, not recalling something from training.\n\n**The transcripts are still there to check.** Every quote an assistant gives you can be opened in Koji and read in its original conversation. This is a thirty second check and it is worth doing before a quote goes into a deck.\n\n**The report's own citations are verifiable by design.** Inside Koji, no claim stands on its own. The evidence panel behind it holds the quotes and the link to the source interview.\n\nTreat an assistant's summary the way you would treat a colleague's: useful, fast, and worth spot checking on anything load bearing.\n\n## Setting Up a Study So the Answers Are Good\n\nThe quality of any answer is set by the interviews, long before anyone asks a question.\n\n### Capture the attributes on the way in\n\nWhatever you want to compare later, collect it before the interview. Import respondents with personalised links carrying role, company, plan tier or region, or ask for it in the screener. A transcript without participant context is readable but not comparable, and comparison is half the value.\n\n### Use structured questions for the numbers\n\nPair open questions with [structured questions](/docs/structured-questions-guide): scale, single_choice, multiple_choice, ranking, yes_no. Then \"nine of twenty three rated onboarding two or below\" is a typed value from the data, not an impression. The open follow-up explains the number, the structured question anchors it.\n\n### Let the quality gate work\n\nThe [quality gate](/docs/how-the-quality-gate-works) scores every interview from 1 to 5. Only interviews scoring 3 or above are billed and carried into the report, so one word answers and abandoned sessions do not quietly become findings.\n\n## A Sample Session\n\nA product team planning a quarter, with the connector set up in their assistant:\n\n> **You:** Read my two pricing studies from this year and tell me the strongest evidence that users would pay for a team plan.\n>\n> **Assistant:** Nine of twenty three participants across the two studies explicitly said they would pay for a team plan. Six of the nine were agencies or small consultancies.\n>\n> - \"I'd pay $50 per seat tomorrow if it had role-based permissions.\" (Agency owner, August study)\n> - \"We're ready to upgrade to a team plan as soon as you have one.\" (Consulting lead, March study)\n>\n> **You:** Now the other side. Who said they would not?\n>\n> **Assistant:** Four participants pushed back. Three said their team was too small to justify per-seat pricing, one wanted a flat rate for unlimited seats.\n>\n> - \"Our team's only three people, per-seat would cost us more than the individual plan.\" (Founder, March study)\n\nEvery claim has a count, every quote comes from a transcript you can open, and the disconfirming side is one follow-up away. Note that the cross study part is what needed the connector. Either study on its own was already answered in its report.\n\n## What This Replaces\n\nTeams that set this up retire a few old habits:\n\n- **The \"let me re-read the transcript\" meeting.** The report's evidence panel puts the quote next to the claim while the meeting is still running.\n- **The \"I'll go look at the spreadsheet\" follow-up.** The responses grid is the spreadsheet, already filled in, and it exports to CSV.\n- **The \"did we ever ask about X?\" thread.** One question to an assistant with the connector.\n- **The manual tagging backlog.** Themes and structured answers are produced when each interview is analysed.\n- **The \"what would users say if we asked?\" hypothetical.** Replaced by what users said when you did.\n\n## When to Read the Report, When to Ask an Assistant\n\nRead the report when the question is about one study, when you need the evidence behind a claim, when you are circulating findings to stakeholders, or when you want the canonical version of what a study found. The report is versioned and its public link is the thing to share.\n\nAsk an assistant when the question spans studies, when you want the answer phrased for a specific audience, when you are drafting a summary or a follow-up study, or when you want to interrogate the data in a way no report section anticipated.\n\nThey are not competing. The report is the record. The assistant is how you interrogate it.\n\n## How to Start\n\n1. **Sign up at koji.so.** Free tier, no credit card.\n2. **Run a study** so you have analysed interviews to work with.\n3. **Read the report** and click a citation. Watch the evidence panel open with the quotes and the link into the transcript. Most questions end here.\n4. **Connect the MCP connector** at `https://www.koji.so/api/mcp/mcp` over OAuth, in Claude, Cursor or your client of choice.\n5. **Ask a disconfirming question first.** It is the pattern that shows the difference between a summary and the evidence fastest.\n\n## Related Resources\n\n- [Koji MCP Integration Overview](/docs/mcp-overview) - Connecting Claude, Cursor or any MCP assistant to your research\n- [Reading Your Research Report](/docs/reading-your-research-report) - The report, the evidence behind each claim, personas and question coverage\n- [Exporting Research Data](/docs/exporting-research-data) - CSV and JSON export when you want the raw data\n- [Structured Questions Guide](/docs/structured-questions-guide) - The question types that give answers real numbers\n- [Understanding Themes and Patterns](/docs/understanding-themes-patterns) - What the analysis extracts from each interview\n- [How the Quality Gate Works](/docs/how-the-quality-gate-works) - Why weak interviews do not become findings\n- [Viewing Interview Transcripts](/docs/viewing-interview-transcripts) - Reading one conversation end to end\n","category":"Reports & Analysis","lastModified":"2026-09-20T18:39:37.446092+00:00","metaTitle":"Ask Questions of Your Interview Transcripts: Koji and MCP | Koji","metaDescription":"Get answers from your interview transcripts with verbatim quotes. The report links every claim to its evidence, and the Koji MCP connector lets Claude or Cursor read your studies and answer across them.","keywords":["chat with interview transcripts ai","query interview transcripts","natural language transcript search","ai chat research data","interview transcript chat","ask questions research data","chat with research repository","llm transcript query","transcript analysis chat"],"aiSummary":"Koji has no chat box over transcripts. The same job is done two ways. Inside Koji, every interview is analysed on its own and the report opens with the finding: key takeaways, themes with evidence counts, a section per question, personas and question coverage. Every claim carries a citation that opens an evidence panel with the supporting quotes and a link into the full transcript, so nothing has to be taken on trust. Outside Koji, the MCP connector at https://www.koji.so/api/mcp/mcp (OAuth, works with Claude, Claude Code, Cursor and any Model Context Protocol client) exposes koji_list_studies, koji_get_study_data, koji_get_interviews and koji_get_transcript, so an assistant can answer free-form questions, compare studies, surface disagreement, or draft a summary from the real conversations. Cross study questions are the main reason to connect it, since a report belongs to one study. Answer quality is set before the interviews run: capture the attributes you want to compare through imported personalised links or a screener, pair open questions with structured questions so numbers come from typed data, and let the quality gate keep interviews scoring below 3 out of the report.","aiPrerequisites":["Have completed or published at least one Koji study","Familiar with basic interview transcript workflows","Comfortable with natural-language search UIs"],"aiLearningOutcomes":["Use the four highest-value chat query patterns (disconfirming, cross-study, persona-specific, prior-art)","Set up your corpus for high-signal chat: participant metadata, structured questions, quality gate","Verify chat answers are grounded with verbatim quotes and participant IDs","Replace re-reading meetings and prior-art Slack threads with single chat queries"],"aiDifficulty":"beginner","aiEstimatedTime":"11 min read"}],"pagination":{"total":1,"returned":1,"offset":0}}