obsei

release

obsei 1.0: private customer feedback in five minutes

obsei 1.0 is out: a self-hosted Voice of Customer pipeline that strips PII before any model sees it, can run its classifier on your machine, and plugs into your AI tools.

By 5 min read

obsei logo with the words: obsei 1.0, private customer feedback in five minutes

Ten days ago I wrote that obsei is back. Today obsei 1.0 is on PyPI, and a plain pip install obsei gives you the new obsei instead of the 2023 library.

obsei is the self-hosted, open source alternative to Enterpret and Unwrap that strips PII before any LLM sees it. It collects reviews, tickets, surveys and community posts; redacts personal data on the way in; labels and groups everything into themes across languages; and sends what matters to Slack, Jira, Linear or your warehouse. It runs on your machine or your servers. Nothing leaves unless you allow it.

Try it

You need uv. No model server, no API key, nothing to export:

mkdir voc && cd voc
uvx --from "obsei[recommended]>=1.0" obsei init
obsei try

obsei init creates a project with sample feedback in ten languages, generates the encryption key for you, and downloads a multilingual model once (about 2.2 GB). obsei try shows what obsei would store, without storing anything:

┏━━━━━━━━┳━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ source ┃ lang ┃ labels                       ┃ text                          ┃
┡━━━━━━━━╇━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ sample │ en   │ bug 0.85 · negative 0.93     │ The app crashes every time I  │
│        │      │                              │ open the camera. Please fix!  │
│ sample │ es   │ praise 0.41 · positive 0.87  │ Me encanta la nueva función   │
│        │      │                              │ de búsqueda.                  │
│ sample │ en   │ complaint 0.39 · neutral     │ Support never replied. Reach  │
│        │      │ 0.75 · review                │ me at <EMAIL>, or I am        │
│        │      │                              │ cancelling.                   │

The email address is gone before the model reads the message. The low-confidence answer is marked review, so a person decides instead of the model guessing.

Then point it at something real. Any public App Store app, GitHub repository, Hacker News thread, WordPress plugin or RSS feed works without a key:

obsei try https://apps.apple.com/us/app/whatsapp-messenger/id310633997
obsei try https://github.com/astral-sh/uv

What’s new since the release candidates

A classifier that runs on your machine. The default model is now bge-m3-zeroshot-v2.0, running inside obsei. I benchmarked six small models on the demo’s 136 synthetic records in 11 languages. It was the most accurate: .82 on intent, .90 on sentiment and .91 on team routing. The decision model I had recommended before, Julia-1, came last at a mean of .69. The full table and its caveats are in the docs. You can still use Ollama, any OpenAI-compatible model or a decision model instead.

Measure it on your own data. obsei eval gold.csv scores the classifier against a labelled sample. For each question it reports accuracy, calibration and how much goes to review, and it suggests a confidence cutoff. obsei correct stores human fixes, and obsei reprocess relabels stored feedback after you change a question or a model.

Set up without editing YAML. obsei add appstore or obsei add slack asks for each field, links to the page where you get the key, keeps secrets out of obsei.yaml, and previews three records. obsei check fetches one record from each source, checks sink credentials where that sends nothing, and says how to fix what fails. obsei integrations lists every source and sink: 22 and 9, counting the first-party plugins. The published JSON Schema gives you autocomplete in VS Code.

Alerts on themes, not just messages. obsei can tell you when a theme spikes against its seven-week baseline, when a new theme appears, when the review queue grows, or when a pipeline fails. It can also send a weekly digest. Each alert carries counts, the baseline and a few redacted quotes, and each has a cooldown, snooze and acknowledge.

Signed webhooks and 100+ notification services. The webhook sink sends CloudEvents signed with Standard Webhooks, retries failed deliveries and logs them. So n8n, Zapier or your own service can trust what arrives. The new notify sink reaches Telegram, Discord, email and about 100 other services through Apprise.

Agents can read, and propose, but not act alone. The MCP server marks customer text as untrusted, so an agent quotes it instead of following instructions hidden inside a review. Agents can now propose a Jira, Linear or GitHub issue. Nothing is created until a person runs obsei approvals approve, and the audit log records who approved it.

Installs for your AI tools. obsei ships as a plugin for Claude Code, Codex, the Copilot CLI and Gemini CLI, installed straight from the GitHub repository; it isn’t in those tools’ official directories yet. Cursor and VS Code get the MCP server from an install button, and the server is listed in the MCP Registry for other clients:

/plugin install obsei --marketplace obsei/obsei      # in Claude Code
gemini extensions install https://github.com/obsei/obsei

The AI tools guide covers the rest, including Cline, Zed, JetBrains and Goose.

What isn’t perfect yet

  • The model makes predictable mistakes. On the demo set its most common errors are praise read as a bug report and negative messages read as neutral. Run obsei eval on a sample of your own feedback before trusting it for routing.
  • Most connectors are beta. Seven sources and sinks are marked verified, which means they have a live test against the real service. The rest are tested against recorded responses only. The sources and sinks pages show which is which.
  • Apple’s public review feed is unreliable. It has returned empty pages for many apps since mid-2026. obsei tries several URL forms, and App Store Connect is the reliable route for your own apps.
  • There is no hosted version. That is deliberate, but it means you run it yourself.

Coming from 0.0.x?

1.0 is a rewrite, and old configurations won’t load. The migration guide maps the old observers, analyzers and informers to the new sources, enrichers and sinks. The 0.0.x code stays on the legacy/0.0.x branch.

What’s next

The roadmap for 1.x starts with model quality on the mistakes above, then adapters that put obsei inside tools people already use: a dlt source and a LangChain document loader. I would rather build what you ask for than guess, so tell me what’s missing.

Lalit Pagaria · LinkedIn · GitHub