CrewForm is an open-source AI agent orchestration platform — think "Slack for AI workers." You create agents, wire them into teams (pipeline, orchestrator, or collaboration mode), and let them handle tasks. BYOK, self-hostable, all three agentic protocols (MCP + A2A + AG-UI).
Today we're shipping v1.8.0 with three major features we're really excited about.
💬 Embeddable Chat Widget
You can now deploy any CrewForm agent as a chat bubble on your website with a single script tag:
<script
src="https://runner.crewform.tech/chat/widget.js"
data-key="cf_chat_your_key_here"
data-theme="dark"
data-position="bottom-right"
async
></script>
That's it. No React, no build step, no npm install. Paste it before closing body tag and your visitors can chat with your agent in real-time with streaming responses.
What we built under the hood:
Standalone Vite build that outputs a single widget.js bundle
Shadow DOM isolation — zero CSS conflicts with your existing site
Domain whitelisting via cf_chat_ API keys
Session memory — conversations persist across page reloads
Rate limiting per visitor
Settings UI with embed snippet generator and live preview
The widget connects directly to your task runner via SSE, so responses stream in word-by-word. We bundle it inside the Docker image so self-hosted deployments get it for free.
🔍 Knowledge Base: Hybrid Search + Retrieval Testing
Our RAG pipeline got a serious upgrade. Previously it was basic vector search — now it's production-grade.
Hybrid Search
We combine vector similarity (cosine distance via pgvector) with PostgreSQL full-text search (tsvector + ts_rank_cd):
Final Score = (0.7 × cosine_similarity) + (0.3 × ts_rank)
The over-fetch + rerank strategy retrieves 2× results from each method, deduplicates, and returns the top-K. This dramatically improves recall for queries that mix semantic meaning with specific keywords.
Metadata Tags
Documents can now be tagged (e.g., "FAQ", "Technical", "Policy") with GIN-indexed filtering. Tag your docs, then filter search results by tag — both in the UI and via the API.
Retrieval Tester
This is my favorite part. Before deploying retrieval to agents, you can test it interactively:
Type a query, see matched chunks with color-coded similarity scores
Toggle between vector-only and hybrid search
Filter by document or tag
Adjust top-K (1–20)
See response times for benchmarking
No more "deploy and pray" — you can validate retrieval quality before it touches your agents.
📦 Agent/Team Export & Import
Share agents and teams as portable JSON files:
{
"format": "crewform-export",
"version": 1,
"type": "team",
"data": {
"name": "Content Pipeline",
"mode": "pipeline",
"agents": [
{ "ref_id": "abc", "role": "researcher", "agent": { ... } },
{ "ref_id": "def", "role": "writer", "agent": { ... } }
]
}
}
Export an agent or team with one click — the JSON includes everything (model, prompt, tools, voice profile, config). Team exports are self-contained with all member agents embedded inline.
Import into any workspace — agent IDs get remapped automatically to maintain referential integrity in team configs. Custom tools are stripped (not portable across workspaces).
This is the foundation for agent sharing — export a team, send it to a colleague, they import it and they're running.
Also in this release
AG-UI Rich Interactions — Agents can now pause mid-execution and ask for approval, data confirmation, or present choices
Marketplace Agent README — Published agents can include rich Markdown documentation
License Key Validation — HMAC-SHA256 cryptographic verification
OpenTelemetry + Langfuse — Opt-in tracing for LLM calls, tool use, and team runs
Try it
🌐 Hosted: crewform.tech — free tier, no credit card
🐳 Self-hosted: docker compose up -d
⭐ GitHub: github.com/CrewForm/crewform
📖 Docs: docs.crewform.tech
💬 Discord: discord.gg/TAFasJCTWs
We'd love feedback — especially on the hybrid search and retrieval tester. What features would you want next?