A local model by default
Labels with confidence from a multilingual model on your CPU, or any OpenAI-compatible LLM or decision model. obsei eval measures it on your data.
New obsei 1.1.0 is out
Open source Voice of Customer · Self-hosted · Apache-2.0
Know what customers are saying. Keep what they said private.
The open source, self-hosted alternative to Enterpret and Unwrap. obsei reads reviews, tickets, calls and surveys in many languages and strips personal data before any model sees it. Nothing leaves your infrastructure unless you allow it.
$ uv tool install "obsei[recommended]"
$ obsei init && obsei tryArrivesApp 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].
- 20sources, plus plugins
- Self-hostedyour servers, your models
- Zerotelemetry
- Apache-2.0no open core
Open source since 2020 · 1,400+ GitHub stars · Joint runner-up, RBI HARBINGER 2021
Why privacy-first feedback analytics
- Feedback is personal dataReviews, tickets and calls carry names, emails, phone numbers and order ids.
- Most tools copy it to their cloudEvery copy is one more breach risk and one more processor to vet.
- obsei keeps it at homeRedacted at ingest, analysed where it lives. Share insights, not identities.
One pipeline, inside your walls
Collect
- App Store
- Google Play
- Zendesk
- Intercom
- Gong
- GitHub
- Bluesky
- SQL
- Webhooks
- + any REST or MCP
Your infrastructure
- RedactPII, names optional
- Classifylocal model or any LLM
- Themeacross languages
- Storeencrypted
Act
- Studio
- AI agents via MCP
- Jira
- Linear
- Warehouse
Your agents answer, with receipts
An MCP server, with plugins for Claude Code, Codex, Copilot and Gemini CLI that install from the GitHub repository. Results carry record ids your agent can cite, never author identities. Customer text is marked untrusted, and the only write, a proposed issue, waits for your approval.
Connect an agent →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 →- model unsureHuman review
- bugconfidence ≥ 0.8urgentJira + on-call
- billingBilling channel
- everything elseWarehouse
Every theme, across languages, in one view
Built for teams that can't ship customer data to someone else's cloud
Alerts on themes
Spikes, new themes, weekly digests and pipeline health, sent to Slack, a signed webhook, or Teams, Telegram and email with obsei[notify].
Setup without YAML
obsei add asks for each field and links to where you get the key; obsei check fetches from each source and checks sink credentials.
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 from fewer than k people are hidden, in Studio, exports and the MCP server.
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.
Themes across languages
The same issue raised in English, Japanese and Portuguese becomes one theme, with trends and near-duplicates.
Plugins
Add a source, enricher or sink as a small Python package.
Frequently asked questions
What is obsei?
obsei is an open source, self-hosted Voice of Customer pipeline. It collects customer feedback from app stores, support desks, communities, files and your own databases, removes personal data as it arrives, labels and groups it into themes across languages, and sends what matters to Slack, Jira, Linear, a warehouse or your AI agents.
Is obsei free?
Yes. obsei is Apache-2.0 licensed, with no paid tier, no open-core split and no hosted version. You run it on your own laptop or servers.
How is obsei different from Enterpret, Unwrap and other feedback analytics tools?
Hosted feedback tools process your customers' text on their servers. obsei runs on your infrastructure, strips PII before any model sees the text and can label feedback with a model on your machine, so customer data doesn't have to leave your network.
Does obsei send customer feedback to an LLM?
Not by default. Redaction runs at ingest, the local classifier runs inside obsei, and the default egress policy blocks model and sink traffic to outside hosts. Feedback is sent to a model or service only if you configure it and allow its host.
Which feedback sources does obsei support?
20 built in: the App Store, App Store Connect, Google Play, Zendesk, Freshdesk, Intercom, Gong, GitHub issues, Hacker News, Bluesky, YouTube, RSS, email (IMAP), SQL databases, CSV and JSON Lines files, file drops, webhooks, any REST API and any MCP tool. Reddit and Typeform source plugins install from the GitHub repository.
Can AI agents like Claude, Cursor or Codex use obsei?
Yes. obsei includes an MCP server, listed in the MCP Registry, and plugins for Claude Code, Codex, the GitHub Copilot CLI and Gemini CLI that install from its GitHub repository. Cursor and VS Code add the server with an install button. Agents search, count and read redacted feedback with record ids to cite. Customer text is marked untrusted, and the only write, a proposed issue, waits for a person to approve it.
How accurate is the classifier?
On obsei's 136 synthetic demo records in 11 languages, the default model, bge-m3-zeroshot-v2.0, scored 0.82 on intent, 0.90 on sentiment and 0.91 on team routing. The expected answers come from the demo's issue definitions, not human review. Run obsei eval on a labelled sample of your own feedback to measure it on your data and pick confidence cutoffs.
How do I install obsei?
Run uv tool install "obsei[recommended]", then obsei init and obsei try in an empty folder; pip install "obsei[recommended]" works too. The first run downloads the local model once (about 1.2 GB) and needs no API keys. The container image ghcr.io/obsei/obsei has the sources and sinks but not the local classifier.
Try it on your machine
Sample feedback in ten languages, no API keys. The first run downloads a local model once (about 1.2 GB).