Dejima
Dejima provides a self-hosted, open-source runtime for securely running fleets of AI coding agents on your own hardware. It isolates each project in a sealed container "island" with brokered, read-only access to host files and MCP servers, while a terminal TUI manages sessions, costs, and audit trails. The platform supports terminal agents like Claude Code and Codex, headless agents, and local models via Ollama.
- Developer Tools
- AI Agents
- Artificial Intelligence
- Software Only
- Cybersecurity
Funding
Founders
Founder details are not available yet.
Product
Problem
Running AI coding agents typically requires either manually wiring together containers, networking, SSH sessions, and credentials—often becoming an infrastructure engineering project—or handing agents broad, uncontrolled access to a host machine. This complexity and security risk make it difficult for individual developers and teams to run agents safely and efficiently on their own hardware or cloud environments.
Solution
Dejima provides a complete, self-hosted runtime that manages the entire lifecycle of AI coding agent fleets. It runs agents inside sealed, per-project "islands" (containers) with isolated git worktrees, granting agents only the host files and MCP servers you explicitly approve, with read-only access by default. A terminal TUI serves as a single dashboard to launch, monitor, hibernate, and resume any number of agents across projects, with session state surviving disconnects. Dejima supports terminal agents like Claude Code and Codex, headless agents like OpenClaw, and local open-weights models via Ollama. A hash-chained audit trail and egress logs provide verifiable proof of agent activity, while an HTTP/WebSocket API and webhooks enable integration and product development.
Target Audience
Dejima targets individual developers and engineering teams who run AI coding agents on their own Mac mini, Linux box, or cloud VPS and need a secure, auditable, and manageable alternative to manual session tooling or managed cloud sandboxes. It also serves organizations with regulated or client-confidential work, where retaining code and data on self-owned hardware is a requirement.
Features
- Sealed container-based "islands" per project, each with its own git worktrees and shared home, isolating agents from the host machine and other projects
- Brokered host access with deny-all-by-default policy, allowing explicit, read-only grants of single host folders
- Brokered MCP servers with per-island grant controls, ensuring agents only access curated tools
- Hash-chained audit trail for every brokered crossing, providing a tamper-evident, verifiable log of agent actions
- Unified TUI for fleet management with per-agent status, CPU/memory, token cost tracking, and one-key session resume across devices
- Built-in agent messaging (`dejima msg`), lifecycle verbs (`hibernate`, `wake`, `reset`, `purge`), and scheduled wakes for automated tasks
- HTTP/WebSocket API and webhook events, alongside a TypeScript-style SDK, for building custom applications and workflows