Berd - Desktop AI Agent Builder for Creative Work | InferenceSaver
Learn how we use Berd, the playful desktop app for building with AI agents. Individual agent workspace, skills system, MCP connections, and handover to Buzz for team collaboration.
Berd is a weird, playful desktop app for building with AI agents. Where Buzz is the team multiplayer workspace - channels, rosters, collaborative agents - Berd is the single-player creative workshop. It's for building hyper-specific agents that work on your projects without stopping, without asking questions, just doing the work step by step until it's done.
We use Berd as the individual creative layer in our agency pipeline. It's where we build and experiment with agents before bringing them into the team workspace through our handover skill. The core philosophy is 'less chatting, more building' - and that's exactly how it feels.
Key capabilities:
- Individual agent workspace - build agents that work on projects autonomously, step by step, without interruptions
- Skills system - install, manage, and create skills that define what your agents can do
- MCP server connections - connect your agents to external tools and data sources
- Multi-provider support - use Claude, GPT, DeepSeek, or any model on the InferenceSaver API - you're not locked into one provider
- Handover to Buzz - build agents individually in Berd, then sync them into the Buzz team workspace
Berd is built for people who want to do interesting things. Artists, designers, builders. People who want to experiment and create without being locked into a corporate workflow. And when you pair it with the InferenceSaver API for cheap model inference, you get an agent-building environment that's both playful and production-ready.
Rendering diagram…
Build agents that just work
Give an agent a project and it goes - it doesn't stop, doesn't ask questions, works through the project step by step until it's done.

Skills make agents capable
Install skills from the library or create your own. Each skill defines what your agent can do - from web research to content creation to code generation.

Handover to the team
Build agents in Berd, then sync them into Buzz via the handover skill. Your personal creations become team assets without rebuilding.

DeepSeek V4 Flash - the affordable engine for Berd agents
Every agent in Berd can be powered by DeepSeek V4 Flash via the InferenceSaver API at $1 for 100 million tokens - the best deal on the internet for agent builders.
| Rank | Model | Capability | Context | Max Output | Input Tokens | Output Tokens | Savings | |
|---|---|---|---|---|---|---|---|---|
| #1 | Deepseek V4 Flash deepseek-v4-flash | Chat | 128K | 16.4K | $0.04$0.02per 1M tokens | $0.04$0.02per 1M tokens | up to 70% off |
How We Use Berd in Our Agency
Berd is the individual creative layer of our agency pipeline. Here's how it fits into the bigger picture.
Building Hyper-Specific Agents
In Berd, we build agents that are designed for one specific job. A competitor research agent that monitors GitHub repos. A content creation agent that builds comparison pages. A social media agent that drafts outreach messages.
Each agent gets:
- A project with clear instructions
- Skills that define its capabilities
- MCP connections to the tools it needs
- A model from the InferenceSaver API - we use DeepSeek V4 Flash for most tasks because it's $1 for 100 million tokens
The agent works autonomously. It doesn't stop to ask for permission. It works through the project step by step, and when it's done, it reports back.
The Handover to Buzz
Once an agent is working well in Berd, we use our handover skill to sync it into Buzz. In Buzz, the agent joins the team workspace where it can collaborate with other agents in dedicated channels.
This two-tier approach gives us the best of both worlds:
- Berd for experimentation - try new agent configurations, test different models, iterate quickly without affecting the team
- Buzz for production - once the agent is proven, bring it into the team workspace where it can handle real signals and outreach
The Skills System
The skills system is the heart of Berd's agent capabilities. Each skill is a package of instructions and tools that an agent can use. You can install skills from the library, create your own, or modify existing ones.
Combine this with the InferenceSaver API and you've got an agent-building environment that's both playful and powerful.
For the full picture of how our agency pipeline works, read about Buzz CLI, Parallel CLI, and Treg.
# Example: running a Berd agent with a skill
berdctl agent run my-agent --skill competitor-research --project "Research AI inference competitors"
# Example: installing a skill
berdctl skill install social-media-outreach
# Example: using InferenceSaver API as the model provider
# Set this in Berd's AI provider settings:
# Provider: InferenceSaver
# API Key:
# Model: DeepSeek V4 FlashWhy Berd Plus InferenceSaver Works
The combination of Berd's creative agent-building environment and InferenceSaver's cheap model inference creates a powerful workflow for anyone building an AI agency:
- Experiment without cost anxiety - DeepSeek V4 Flash at $1 for 100 million tokens means you can iterate on agent designs without worrying about API costs
- Multi-model flexibility - use the best model for each job - Claude for complex reasoning, DeepSeek for high-volume tasks, GPT for creative work
- Open source foundation - Berd is built on open-source tools, Buzz is open source, the InferenceSaver API gives you access to open models at the best prices
- From solo to team - start in Berd, scale to Buzz, use Twenty CRM as the backend switchboard
The future of building is like playing a video game. You're the player. The agents are your party. The tools are your inventory. And the pipeline is your quest.
Get started with InferenceSaver today and build your first Berd agent in minutes.
Ready to build your own agents?
InferenceSaver gives you the model inference layer. Berd gives you the creative agent-building environment. Together, they're the foundation for your AI agency.
Berd is a desktop app by Block. InferenceSaver provides the model inference layer with DeepSeek V4 Flash at $1 for 100 million tokens.