SWIRL 5 is here: one canonical answer, from everywhere you keep knowledge.

The private knowledge layer for enterprise AI - now with a first-class MCP server, Slack access, one-click Railway deploys, and more.

Hi - it's been a big summer for SWIRL. On July 15 we shipped SWIRL 5, and since then we've been rolling out the pieces that make it useful in real workflows: an MCP server your agents can call directly, a Slack integration, a one-click Railway deploy, and the feature at the center of it all - the canonical document finder.

A year ago, "federated search across your systems, then re-rank the results" was a product you could sell. Today it's table stakes - everything finds the document. The interesting question moved: which of the versions you found is the one your organization actually stands behind? That's what this release is built around.

🚀 Meet SWIRL 5

SWIRL 5 is the private knowledge layer for enterprise AI. It runs federated AI search and RAG across 100+ systems in real time - without copying, indexing, or moving a single byte - and serves ranked, permissioned, organization-approved answers to people and to any AI agent.

Under the hood: a local three-pass relevancy pipeline (BM25 → embedding re-rank via reciprocal rank fusion → cross-encoder re-rank), bounded prompt assembly (~3,000 tokens, max 10 sources), and hallucination checks that flag when a model's answer strays from its citations. Plus a Topic Dashboard, streaming search, multi-model support, and workspaces.

Read the SWIRL 5 launch post →  |  Full release notes →

🎯 The canonical document finder

This is the headline feature. When SWIRL finds nine versions of "the data retention policy" scattered across OneDrive, SharePoint, and email, it clusters near-duplicate versions by embedding similarity, scores each on source authority, naming convention, recency, and format, and elects a canonical version automatically. A person can override that election with one click - and that pin isn't just a UI bookmark, it's a retrieval-layer override: every downstream consumer, including MCP agents, gets that single source of truth from then on.

The same semantic cache that powers version election also does query-level caching keyed by embedding rather than exact string match, so repeat questions get faster, cached answers grounded in the pinned document. Data never moves.

Watch: "One Search, One Canonical Answer" →

🔌 A first-class MCP server

SWIRL 5 exposes its search and RAG stack to any MCP-speaking agent via search, search_rag, read_document, score_document, list_providers, and chat tools - dev/demo auth over a token, production auth via OAuth 2.1 with PKCE and per-caller JWT-to-user mapping, so agent results stay permission-trimmed per person.

In our CLI demo, Claude searched the enterprise over MCP and - unprompted - flagged that three versions of a policy were in circulation across two systems. Agents with real retrieval notice real problems.

We also shipped a free, Apache-2.0 Claude Code plugin with eight guided slash-commands (/swirl:install, /swirl:provider, /swirl:connector, /swirl:rag, /swirl:mcp, /swirl:troubleshoot, and more) that take a fresh install through connectors, RAG, and MCP wiring in about an hour instead of a day. Install with:

claude plugin marketplace add swirlai/swirl-claude-plugin
claude plugin install swirl@swirl --scope user

Announcing the Claude Code plugin →  |  Watch: Claude + SWIRL over MCP →

💬 SWIRL, right inside Slack

We turned Slackbot into an enterprise research agent over MCP. Ask Slack "search OneDrive for our insurance policy" and it comes back with coverage limits in a table, flags exclusions, and links straight to the exact sentence in SharePoint via a text-fragment deep link - not just the document, the passage.

Two lessons from building it, for the technically curious: Slack's MCP tool-call timeout is tighter than SWIRL's RAG polling window, so we tuned SWIRL_MCP_RAG_POLL_TIMEOUT; and MCP tool descriptions turned out to be UX - they're what let the bot accurately describe its own capabilities to whoever's asking. Bottom line: agents don't need your data ingested into them, they need a governed search layer.

🚂 Self-deploy on Railway

We stopped fighting Azure Marketplace and started shipping on Railway. SWIRL Enterprise is now a one-click Railway template - request a license, paste it into Railway, add an LLM API key, connect your sources, and search. It runs entirely in your own cloud workspace: SOC 2 Type II, SOC 3, and HIPAA support at the enterprise tier, per-second billing, and headroom up to 1TB RAM / 1,000 vCPU. For a small team, all-in cost typically lands in the low hundreds of dollars a month.

Deploy SWIRL Enterprise on Railway →

⚖️ On the horizon: MikeOSS

MikeOSS, the open-source legal AI front-end, will soon support MCP - which means it'll be able to plug directly into SWIRL's federated search and grounded RAG as a backend, the same way Claude and Slack do today. More details as that lands; we'll cover it in a future update.

📎 Also this summer

SWIRL Community is Apache 2.0 and open source on GitHub.

You're receiving this because you follow SWIRL's work. SWIRL was created by Sid Probstein and released under the Apache license.

← Back to swirlaiconnect.com