Skip to content

MCP and agents

obsei mcp serves these tools over stdio:

Tool Use
feedback_stats Counts and average rating grouped by source, instance, sentiment, intent, lang (language), rating, route, day, week or month
search_feedback Matching feedback, newest first, as citable evidence
get_feedback One record by id, to verify a citation
list_themes Stable themes with trends (rising as Studio shows it), sources, languages and intents (k-anonymous)
propose_issue Queue an issue for a person to approve (only when a GitHub, Jira or Linear sink is configured)

Results contain redacted text, labels, source, time and record id. They never contain author pseudonyms. The server also offers a voc_report prompt. Its initialize handshake negotiates MCP protocol 2024-11-05 to 2025-11-25 (a client offering 2026-07-28 gets 2025-11-25); the bundled MCP SDK also answers 2026-07-28 requests that carry their version in _meta. Every tool declares readOnlyHint, destructiveHint and openWorldHint: the four read tools are read-only, and propose_issue writes only to obsei’s own queue.

Each evidence item’s text is what a customer wrote, and anyone can write a review. Items carry "untrusted": true, and the tool descriptions and server instructions tell agents to quote that text and never follow instructions in it. Theme labels and descriptions are written by a model from that text, so treat them the same way.

An agent cannot act outside obsei. propose_issue(target, title, body, record_ids) stores a pending approval and returns its id; nothing is sent. target is the key of a github_issues, jira or linear sink in obsei.yaml (the tool description lists them; a key that names different sinks in several pipelines becomes pipeline/key). Cited record ids must exist.

A person then decides, with the CLI or, as admin, over HTTP:

Terminal window
obsei approvals list --status pending
obsei approvals show apr_... # the proposal and the exact issue text
obsei approvals approve apr_... --by ana
obsei approvals reject apr_... --by ana --reason "duplicate of #12"

Approving opens the issue with that sink’s settings. The issue starts with “Proposed by an AI agent via obsei; review before acting” and lists the cited records. An approval is carried out once: approving it again is refused, and if opening the issue failed you can approve it again without risking a second issue, because the sink first looks for one tagged with the approval id. Every proposal and decision is written to the audit log.

Over stdio, obsei mcp opens the store read-only, and read-write only for the moment it stores a proposal. DuckDB allows one writer, so while obsei serve or obsei run holds the store, send proposals through obsei serve’s /mcp instead.

The rule behind this: fixed labels from a classifier may pick routes on their own, but any free text an agent writes after reading feedback goes through a person, because private data, untrusted content and a way to send messages out together are what a prompt injection needs.

Terminal window
/plugin marketplace add obsei/obsei
/plugin install obsei@obsei

This installs from obsei’s own marketplace in the GitHub repository; obsei isn’t in Claude’s official plugin directory yet. The plugin adds the MCP server and a voice-of-customer skill. Codex, Copilot, VS Code, Cursor, Gemini and other tools: see Use obsei from your AI tool.

{
"mcpServers": {
"obsei": {
"command": "uvx",
"args": ["--from", "obsei>=1.0", "obsei", "mcp"],
"env": { "OBSEI_DB": "/path/to/obsei.duckdb", "OBSEI_DB_KEY": "..." }
}
}
}

obsei serve exposes the same tools at /mcp (streamable HTTP). Set OBSEI_API_TOKEN; clients send Authorization: Bearer <token>. Put it behind your SSO proxy for per-user access. Give agents an analyst token: proposals then record who made them (agent:<user>), and only admins can approve them.