obsei
New 1.0 release candidate is out

Open source · Self-hosted · Apache-2.0

Voice of Customer without violating privacy

Understand feedback from every channel and language. Personal data is redacted before anything is stored, and nothing leaves your infrastructure unless you allow it.

Get started Live demo 
$ uv tool install "obsei[mcp]>=1.0.0rc1"
$ obsei init && obsei try
One review, before and after obsei

ArrivesApp Store · ★★ · en

Hi, I'm Priya Sharma. Charged twice for order 4471. Call me on +91 98765 43210 or priya@mail.com

Storedauthor psn_7f3a…

Hi, I'm <PERSON>. Charged twice for order 4471. Call me on <PHONE> or <EMAIL>

  • negative
  • complaint
  • theme: Charged twice

Names use the optional local model, obsei[names].

Open source since 2020 · 1,400+ GitHub stars · Joint runner-up, RBI HARBINGER 2021

Why privacy-first feedback analytics

  1. Feedback is personal dataReviews, tickets and calls carry names, emails, phone numbers and order ids.
  2. Most tools copy it to their cloudEvery copy is one more breach risk and one more processor to vet.
  3. obsei keeps it at homeRedacted at ingest, analysed where it lives. Share insights, not identities.

How the controls map to GDPR, DPDP, LGPD, CCPA and more →

One pipeline, inside your walls

Collect

  • App Store
  • Google Play
  • Zendesk
  • Intercom
  • Gong
  • GitHub
  • Bluesky
  • SQL
  • Webhooks
  • + any REST or MCP

Your infrastructure

  1. RedactPII, names optional
  2. Classifydecision model or LLM
  3. Themeacross languages
  4. Storeencrypted

Act

  • Studio
  • AI agents via MCP
  • Slack
  • Jira
  • Linear
  • Warehouse

Your agents answer, with receipts

A read-only MCP server for Claude, Cursor, VS Code and your own agents. Results carry record ids your agent can cite, never author identities.

Connect an agent →
Illustrative conversation

Why are Japanese users unhappy this week?

Mostly login failing after the 2.3 update: 14 reviews, up from 2. Double charges are second.

  • rec_3f9a
  • rec_b210
  • rec_77c4

Act on it, when the model is sure

Ordered routes on labels and confidence. Unsure answers go to a person, not to your issue tracker.

See routing →
Example routes
  1. model unsureHuman review
  2. bugconfidence ≥ 0.8urgentJira + on-call
  3. billingBilling channel
  4. everything elseWarehouse

Every theme, every language, one view

obsei Studio showing themes, a knowledge graph of sources, languages and intents, and feedback in eleven languages
Open the live demo · synthetic feedback in eleven languages

Built for teams that can't ship customer data to someone else's cloud

Decision models or any LLM

Fast labels with probabilities from a small model on CPU, or any OpenAI-compatible LLM.

Privacy by region

Emails, phones, cards and IBANs, plus national IDs from Aadhaar and PAN to SSN, CPF and NHS, checksum-validated where the format allows.

k-anonymous views

Themes and groups from fewer than k people are hidden.

Erasure and audit

Forget by author, source or age, export, and an audit log.

Enterprise access

Roles, sign-in through an SSO proxy such as oauth2-proxy or Cloudflare Access, signed webhooks and a container image.

Plugins

Add a source, enricher or sink as a small Python package.

Try it in five minutes

Sample feedback in ten languages. No API keys.

Read the quickstart Star on GitHub