Dedalus Labs review

Dedalus Labs provides persistent cloud Linux machines for AI agents, plus tools for connecting models to MCP servers. The machines are isolated VMs with their own kernel and root access. Files and installed packages persist when a machine sleeps, so an agent can return to its environment instead of rebuilding it each time.

Machines autosleep after five idle minutes by default and can wake through an API command or SSH connection. Compute is billed while a machine is active, not while it sleeps; storage is billed separately by GiB-hour. Dedalus also lists API-based CPU and RAM resizing without a restart, headless command execution, and interactive SSH access with a PTY.

The MCP gateway supports vendor-agnostic handoffs, tool chaining and streaming. Client tools and SDKs are open source, but the VM infrastructure is proprietary and cannot be self-hosted. That makes Dedalus a managed infrastructure choice, not an open-source way to run your own agent machines.

The main advantage is that the machine is a persistent environment, not a disposable command runner. An agent can retain files and packages across sleep cycles, while autosleep can stop compute charges when the machine is idle. The listed five-minute default and wake options via API or SSH make the operating model clear.

The controls cover more than just starting and stopping a VM. Dedalus lists dynamic CPU and RAM changes through an API call, headless executions, and interactive SSH with a PTY. Its MCP gateway adds a way to chain tools and hand off between providers without tying the gateway to one model vendor.

The entry plans publish useful capacity limits. Hobby is free and allows up to five machines, each with up to 4 vCPU, 16 GiB RAM and 10 GiB storage. Pro is $20 per month, allows up to 20 machines with up to 16 vCPU and 64 GiB RAM each, and includes $20 in monthly compute credits that do not roll over. Those figures make it easier to check whether the service fits before committing.

The listed plan price is only part of the bill. The pricing page lists compute at $0.04536 per vCPU-hour and $0.01458 per GiB-hour of memory, plus $0.0001 per GiB-hour for additional storage. Pro’s $20 monthly compute credits do not roll over, so sustained workloads can run past the subscription price. The page gives no annual Pro price.

Hobby has firm ceilings: five machines, 4 vCPU and 16 GiB RAM per machine, and 10 GiB of storage with no expansion. Pro raises the caps to 20 machines, 16 vCPU and 64 GiB RAM per machine, and 20 GiB of included storage with unlimited expansion—but additional storage is billed. Enterprise pricing is custom, so teams needing its dedicated fleet, SSO, RBAC, audit logs or SLA must request a quote.

There is also vendor lock-in at the infrastructure layer: the VMs cannot be self-hosted. The published Reddit results offer little independent user feedback; the launch thread has no comments, and the other posts are hiring listings.

Dedalus Labs has a free Hobby plan and a $20-per-month Pro plan, according to its pricing page. That is a jump from $0 to $20 monthly, but Pro includes $20 in monthly compute credits; they do not roll over. Hobby instead gets a one-time $20 sign-up credit.

Hobby allows up to five machines, each with up to 4 vCPU, 16 GiB RAM and 10 GiB of storage. That storage cannot be expanded. Pro allows up to 20 machines, 16 vCPU and 64 GiB RAM per machine, plus 20 GiB of storage per machine with unlimited expansion. Enterprise pricing is custom.

Credits do not make usage unlimited. Active compute costs $0.04536 per vCPU-hour and $0.01458 per GiB-hour of memory; additional storage costs $0.0001 per GiB-hour. Compute is billed per second while machines run, and storage per GiB-hour. Machines autosleep after five idle minutes by default, but check how your workload affects active time. The pricing page lists monthly Pro pricing and no annual option.

Moving an agent workload means adapting it to Dedalus’s Linux VMs and APIs. The machines retain files and packages across sleep cycles, and the docs describe command execution and SSH access. If you leave, expect to plan for moving your machine files, packages and task setup to another environment; the published materials do not publish a data-export process or migration tooling.

The client tools and SDKs are open source, but the VM infrastructure is proprietary and cannot be self-hosted. That is the central lock-in: you can inspect or adapt client-side code, but you cannot take Dedalus’s managed VM layer with you. Before switching, check that your agent can run elsewhere and confirm how you will retrieve its data.

We don’t have Reddit posts or comments in the published material, so we can’t substantiate specific reasons users leave. The documented constraints to weigh are proprietary infrastructure, non-expandable Hobby storage and usage-based compute and storage charges.