obsei

obsei is an open-source, self-hosted tool for collecting and analysing customer feedback. It pulls in reviews, tickets, surveys and community posts in many languages, redacts personal data before anything is stored, labels it and groups it into themes, all in an encrypted database on your own machine or servers. Its MCP server lets your agent answer questions about that feedback with quotes you can check. What you can do with it - Ask what customers are talking about and see the recurring themes, with which ones are rising - Size a question with counts and average rating by source, sentiment, intent, language, rating, day, week or month - Pull evidence: matching feedback, newest first, quoted in its own language with a record id for each quote - Check a citation by fetching the record it came from - Propose a GitHub, Jira or Linear issue backed by the feedback, which waits for a person to approve it Where the feedback comes from You set up pipelines with the obsei command line (obsei init, obsei add, obsei run). Sources include App Store and Google Play reviews, Zendesk, Freshdesk, Intercom, Gong, GitHub issues, Hacker News, Bluesky, YouTube, RSS, email inboxes, SQL databases, CSV files and REST APIs. Labels come from a classifier that runs on your machine, or from a model you choose. How it connects The MCP server runs on your machine with uv: uvx --from "obsei>=1.0" obsei mcp. It reads obsei.yaml in the folder it starts in, or the file named by OBSEI_CONFIG (use an absolute path). Set OBSEI_DB for a database kept elsewhere, and OBSEI_DB_KEY when the key isn't in the .obsei/secrets.env file that obsei init writes beside obsei.yaml. The first start downloads obsei and can take a minute. obsei also offers a Claude Code plugin and a Gemini CLI extension that add a voice-of-customer skill, and obsei serve offers the same tools over HTTP behind a token. Good to know - Set up obsei and run a pipeline first: the server answers from feedback obsei has already stored. - The four read tools change nothing. propose_issue only queues a proposal inside obsei; a person approves it with obsei approvals before any issue is opened. - Emails, phone numbers, cards, IBANs and national IDs are redacted and authors are pseudonymised at ingest, and results never include the pseudonyms. - Feedback text is marked untrusted, and the server tells agents to quote it and never follow instructions inside it. - Open source under Apache-2.0, with no telemetry.

• “What are customers talking about this month?”

  1. list_themes
  2. feedback_stats
  Copy

• “How has sentiment changed week by week, by source?”

  1. feedback_stats
  Copy

• “Find complaints about login, in every language, with quotes.”

  1. search_feedback