Block, the company behind Square, Cash App, and Tidal, has released Berd as open-source software under the Apache 2.0 license. The desktop application provides a unified workspace for interacting with AI agents across different models and tools.
Berd is available for macOS, Windows, and Linux, with version 0.6.2 published on GitHub on August 18. The project includes 91 contributors and serves as a daily work surface for tasks like starting chats, attaching files, managing projects, and configuring AI providers. Unlike browser-based tools, it integrates directly with local files, repositories, and agents running on a user’s computer.
A workspace built for visibility and control
Block designed Berd to prioritize clarity over convenience. The interface displays operational details—active projects, files, agents, models, and session states—without hiding them behind a simplified chatbot interface. Failures, unavailable providers, and loading states appear plainly, while configuration settings remain part of the workflow rather than buried in administrative menus.
The application originated from an internal need at Block. Employees used agents like the company’s own Goose, Anthropic’s Claude Code, and OpenAI’s Codex, but the experience was fragmented. “We had capable agents, but switching between them required managing different interfaces and configuration systems,” said Lucinda Bell, Technology Communications at Block. Berd was developed to combine those experiences into one consistent desktop application.
For enterprises, the appeal lies in persistent projects. Users can return to saved collections of files, instructions, and agent configurations without rebuilding context for each task. The design also aims to make agent-based work accessible beyond engineering teams, letting users begin with a conversation and add structure as needed.
Agents with identities, not just chat windows
Berd approaches agent differentiation differently. Instead of treating every agent as a blank chat window, it assigns them roles, instructions, skills, and distinct visual identities. Block created animated characters called “Gloopies” to represent different agent personas—Berdy, Pushback, Choosey, Copycat, Tinker, and Wildcard—each tailored to specific tasks. Pushback, for example, challenges drafts, while Choosey helps narrow decisions.
The visual layer serves a functional purpose. Block’s design documentation explains the avatars make agents recognizable, while their roles and tools make them useful. As users manage multiple specialized agents, the distinct identities act as shorthand for their instructions and permissions, which could become more important as agents handle increasingly complex tasks.
How Berd fits into the agent workspace market
The agent workspace market includes products like OpenAI’s Codex, Anthropic’s Claude Code, and Cursor, all competing for attention. Berd stands out with its open-source license, support for multiple providers, and focus on non-engineering use cases. Unlike most competitors, it isn’t tied to a single vendor’s models or priced as a subscription.
Related: Asana creates secure AI workplace agents
Here’s how it compares:
- Berd: Free and open-source (Apache 2.0). Supports any provider the user configures. Designed for both engineers and non-technical users.
- Codex app: Included with paid ChatGPT plans ($20–$200/month). OpenAI models only. Acts as a command center for coding agents, running multiple agents in parallel.
- Claude Code: Included with paid Claude plans ($20–$200/month) or via API. Anthropic’s Claude models only. Focuses on autonomous coding with deep integration into codebases and terminals.
- Cursor: Pro plan at $20/month, Ultra at $200/month. Supports multiple frontier models. Functions as an AI-native code editor with local-to-cloud agent handoffs and multi-repository workflows.
- Goose: Free and open-source (Apache 2.0). Model-agnostic, including local models via tools like Ollama. The underlying framework for Berd, available as a CLI or desktop app.
Portability matters in Berd’s design. Folders, Git repositories, and plain-text instructions remain accessible outside the application. Skills are file-based, and agents can be exported and shared with their instructions and settings intact.
Brad Axen, Head of AI Capabilities at Block, described Berd as an orchestration layer rather than a new model or runtime. “Agent harnesses like Goose, Claude Code, and Codex manage the loop between a model, its context, and its tools,” he said. “Berd provides the consistent desktop environment around those harnesses.”
Data, telemetry, and enterprise considerations
Berd follows a local-first data model. Conversation history stays on the user’s device in a local session database, while credentials are stored in the operating system keychain. However, when a user sends a prompt to a configured model, the relevant data is transmitted to that model’s provider under the provider’s terms.
Telemetry is disabled by default in official Berd distributions. If enabled, Block collects a random installation identifier and limited usage events, excluding prompts, messages, files, source code, credentials, names, and email addresses. IP addresses are processed for transmission and security but not included in the telemetry payload. Users can disable telemetry in settings, though third-party forks may modify or redirect telemetry behavior.
For enterprises, Berd offers distribution seams—points where organizations can overlay managed provider settings, private agents, and runtime configurations without modifying the public source tree. Block’s internal deployment uses approved devices, model-provider infrastructure, and governed tool connections. “The application can remain open and inspectable, while an organization decides which models, tools, permissions, and policies fit its environment,” Axen said.
Berd isn’t a complete enterprise management solution. The documentation lacks details on features like single sign-on, centrally enforced policies, or audit logs. It functions as endpoint software installed on individual computers, not a centrally hosted SaaS product.
Pricing, models, and the road ahead
Block hasn’t announced a subscription price for Berd, and none appears planned. The economic model relies on users paying for their own model or API costs. The approach emphasizes flexibility: users can select the right model for each task rather than defaulting to an expensive option for every job. The project’s website notes that some tasks require advanced models, while most do not.
Related: Target’s AI edge lies beyond the models
The announcement and GitHub README don’t provide a definitive list of supported model providers. Internally, Block uses a mix of frontier and open-weight models, with the catalog evolving as models improve. “The best choice depends on the work,” Axen said.
While Berd’s source is public, Block doesn’t accept outside pull requests. The README states the project is maintained by a small team at Block, and external contributions are automatically closed. Users can inspect, build, and fork the code, but Block retains control over the upstream repository. For companies considering Berd as infrastructure, this makes long-term maintenance and support plans important.
Block isn’t treating Berd as a revenue product yet. Axen said the immediate benefit comes from internal use—making agent work more consistent while learning how employees interact with the software. Open-sourcing it allows outsiders to inspect, adapt, and provide feedback. A commercial layer around enterprise deployment may follow, but Block hasn’t shared details.
The company’s other AI collaboration tool, Buzz, offers a different approach. Built on the decentralized Nostr protocol, Buzz is a free, open-source workspace where humans and agents collaborate in shared rooms. Unlike Berd’s local focus, Buzz supports multiplayer interactions, with agents holding their own cryptographic identities and permissions. Axen described the split as a matter of trust boundaries: local agents access a user’s filesystem and credentials, while agents in Buzz operate within scoped permissions.
Block’s roadmap blurs that distinction. “Going forward, our focus is Buzz, and we plan to bring the best parts of Berd’s single-player experience into it,” Axen said. The goal is seamless continuity between private and collaborative work, with cross-device support including mobile. For now, Berd remains available as an open-source desktop application, but Block encourages users to adopt Buzz instead.
For IT and security teams evaluating Berd, the key points are clear. Conversation history stays local by default, credentials use the OS keychain, and context sent to models leaves the machine under the provider’s terms. Official telemetry is off by default and excludes sensitive data. Organizations can package their own governed distributions, though central management tooling isn’t yet documented.
The broader implication is that models and agents are becoming interchangeable components in larger workflows. Berd’s approach suggests the surrounding context—projects, skills, tools, and persistent environments—will become the stable layer, not the agents themselves.
