macOS workspace for coordinating teams of AI coding agents with persistent identities, communication, task management, status reporting and shared project context.
Pentagon Studio is a free download; Enterprise deployments are sales-led with no public numeric price verified.
Free access
YES
Platforms
macOS
API
Not confirmed
Open source
NO
Integrations
Claude Code CLI, OpenAI Codex, Gmail, Slack, GitHub
Last updated
2026-09-06
Product sources: Pentagon official product/docs + Y Combinator P26 company page
Pentagon review
Updated 2026-09-08
Product & capabilities
Pentagon Studio is a macOS workspace for coordinating teams of AI coding agents rather than a replacement model or coding runtime. The product combines a spatial workspace with persistent agent identities, communication, teams and channels, tasks, status reporting and shared project context.12
The current docs support Claude Code and OpenAI Codex as agent runtimes and list Gmail, Slack and GitHub connectors.2 This makes Pentagon most relevant as an orchestration layer around existing coding-agent workflows. Hardware-level VM sandboxing is explicitly marked as coming soon on the current product surface, so it is not counted here as an available security capability.1
Pricing & access
Pentagon Studio is currently offered as a free macOS download.1 Enterprise access is sales-led rather than publicly priced, with hosted, VPC and on-premises deployment options described for organizational buyers.1
The free entry point is strong for teams that already have compatible agent runtimes and a Mac-based workflow. The trade-off is narrower platform accessibility: AiToolMap verified a macOS Studio surface, not equivalent Windows or Linux desktop availability, and no public standalone Pentagon API was confirmed in this review.
Evidence & trust
Pentagon publishes unusually specific data-handling boundaries. Its Privacy Policy and Terms state that agent messages, reports, memory, task history and workspace data are stored in the cloud, while source code, documents and files are not stored on Arlington Labs servers; messages may be relayed to AI model providers, and account data is retained while the account is active with deletion described within 30 days after account deletion.34
The company also states controls such as TLS, encryption at rest, audits and access controls, but these remain vendor security statements rather than independent assurance.3 AiToolMap performed no penetration test or source-code audit and did not use the product hands-on. The fixed 50-source review panel produced no eligible exact-product independent review, benchmark or numeric user-rating corpus, so external validation is currently limited and missing evidence does not reduce the score.
Who it's for
Pentagon is best suited to software teams, agent developers and engineering leads who are already using multiple coding agents and need a durable coordination layer around them. Persistent identities, shared tasks/status, agent communication and existing-runtime support are more differentiated for multi-agent teams than for a solo developer who only needs one coding assistant.12
Y Combinator independently corroborates Pentagon as an active Arlington Labs product focused on managing teams of coding agents.5 Teams with strict data-residency or execution-isolation requirements should review the Enterprise deployment options and current legal terms carefully, and should not assume the future VM-sandboxing feature is already available.
Strengths & weaknesses
Strengths
a clear multi-agent coordination surface; persistent agent identities and shared context; task, status, team and channel primitives; Claude Code and Codex runtime support; useful Gmail, Slack and GitHub connectors; a free Studio entry point; and explicit disclosure of what collaboration data is and is not stored in the vendor cloud.123
Weaknesses
early-stage maturity; macOS-only verified Studio access; no eligible independent numeric evidence from the fixed review panel; important agent-message/memory/workspace metadata is cloud-stored; vendor security claims are not independently validated; and hardware-level VM sandboxing remains future-facing rather than a present capability.13