Vibe-Search

Overview

Vibe-Search is a natural-language-first search engine for AI agent outputs, launched October 1, 2026 (92 HN points) as part of the emerging "search-for-agents" category. Where traditional search is for human users querying documents, Vibe-Search is for agents (and their human operators) querying the corpus of agent-generated outputs — emails, files, research notes, code reviews, and decisions that accumulate across an organization's agent fleet.

The thesis is that as agent fleets grow, the corpus of agent outputs becomes larger than the human operators can read. Vibe-Search gives operators a single natural-language query interface across that corpus, with the search happening at the semantic and structured layers rather than the keyword layer.

Vibe-Search ships as an MCP 2.2 server, so any agent can invoke it as a tool without framework-specific code.

Features

• Natural-language queries — type in plain English; no query language to learn
• Semantic + structured hybrid — vector search over embeddings combined with filtered structured queries
• Cross-agent output indexing — index outputs from any agent fleet: emails, files, tickets, code reviews, decisions
• Agent output provenance — every result tagged with which agent wrote it and when
• Time-range filtering — "what did the sales team decide last quarter?" filters by agent + time
• Streaming results — results arrive incrementally as the search progresses, per MCP 2.2 streaming
• MCP 2.2 native — server supports cross-server chaining for multi-source queries
• Framework-agnostic — works with LangGraph, CrewAI, OpenAI Agents SDK, PydanticAI, AgentHub IDE
• Embedding-model pluggable — OpenAI, Cohere, Voyage, or a local embedding model
• Access control — ACLs per agent identity and per user; queries are filtered before ranking
• Audit logging — every query logged with the querying agent's identity

Installation

# Docker
docker pull vibesearch/server:latest
docker run -p 8899:8899 vibesearch/server:latest

# Or via npm
npm install -g @vibesearch/server

# Python SDK
pip install vibesearch

Quick Start

from vibesearch import VibeSearch

vs = VibeSearch.from_env()  # reads VIBESEARCH_API_KEY

results = vs.search(
    "What decisions did the product team make about pricing last quarter?",
    agent_ids=["agent:product-planner", "agent:pricing-analyst"],
    time_range="2026-07-01 to 2026-09-30",
    top_k=20,
)

for r in results:
    print(f"[{r.agent_id} @ {r.timestamp}] {r.text[:200]}")

Core Concepts

Natural-Language Query Parser

Vibe-Search parses free-form natural language queries into a hybrid search plan:

• Intent extraction — what's the user looking for (a decision, a fact, a person, a code change)?
• Time and scope parsing — "last quarter", "in the sales team", "for project X" become filters
• Embedding generation — the query is embedded for semantic search
• Filter composition — agent IDs, time ranges, and labels become structured filters
• Hybrid ranking — semantic similarity and structural relevance combined via Reciprocal Rank Fusion

The parser is powered by a small Mistral Medium model running locally, so the parsing is fast and doesn't add a round-trip to a hosted LLM.

Cross-Agent Output Indexing

Vibe-Search indexes outputs from any agent that produces structured records. The ingestion layer accepts:

• Emails — from Gmail, Outlook, or any IMAP source
• Files — Markdown, PDF, DOCX, code files
• Tickets — from Jira, Linear, GitHub Issues
• Code reviews — from GitHub PRs, GitLab MRs
• Meeting notes — from Notion, Google Docs, or agent-generated notes
• Decisions — structured ADR-style records
• Chat messages — from Slack, Discord, or Teams

Each output is stored with provenance metadata (which agent wrote it, when, under what identity) and indexed both semantically and structurally.

Access Control

Queries are filtered before ranking. If the querying agent doesn't have ACL access to certain outputs, those outputs are invisible to the query — not just deprioritized. The filtering happens at the index level, so a non-privileged agent literally cannot retrieve the outputs.

MCP Configuration

{
  "mcpServers": {
    "vibesearch": {
      "command": "vibesearch-mcp",
      "args": ["--url", "http://localhost:8899"],
      "env": {
        "VIBESEARCH_API_KEY": "***"
      }
    }
  }
}

Exposed tools: vibesearch:search, vibesearch:ingest, vibesearch:forget, vibesearch:query_stats.

Advanced Features

• Federated search — search across multiple Vibe-Search instances without centralizing data
• Result explanations — every result includes why it matched ("matched on agent: pricing-analyst, time: within range, semantic: 0.87")
• Follow-up queries — re-rank results based on the previous query context
• Custom labels — tag outputs with project, team, or domain labels for filtering
• Anomaly detection — surfaces unusual outputs (e.g., an agent suddenly producing outputs in a new category)
• Agent output diff — compare what agent A said about a topic vs what agent B said

Examples

A search across the entire organization:

results = vs.search(
    "Which agents have flagged the auth module as having a security issue?",
    agent_ids=None,  # all agents
    time_range="past 30 days",
)

# Group by agent
from collections import defaultdict
by_agent = defaultdict(list)
for r in results:
    by_agent[r.agent_id].append(r)

for agent, records in by_agent.items():
    print(f"{agent}: {len(records)} hits")

A cross-agent coordination query:

# What did the coder say about the bug the planner raised?
coder = vs.search("auth race condition fix", agent_ids=["agent:coder-v2"])
planner = vs.search("auth race condition", agent_ids=["agent:planner-v1"])
# Cross-reference and surface disagreements

Pros

• First search engine designed for agent outputs, not human documents
• Natural-language queries — no query language to learn
• Provenance tracking on every result
• ACL filtering at the index level for real access control
• MCP 2.2 native for agent integration
• Local deployment option for on-prem requirements

Cons

• Smaller community than general-purpose search tools
• Parser quality depends on the underlying Mistral model
• Index building for large corpora can take significant time
• Access control setup requires careful ACL design

When to Use

• Large agent fleets where operators can't manually read all outputs
• Compliance-heavy environments needing audit trails on agent decisions
• Multi-team organizations where agent outputs span project boundaries
• Teams using MCP 2.2 that want a search primitive

Resources

• Website: https://vibesearch.dev
• GitHub: https://github.com/vibesearch/server
• Docs: https://vibesearch.dev/docs