Devin is a factory you rent. engrams is one you own.
Devin and Factory run the agent, the machines, and your credentials on their platform. engrams runs the agent you choose, in microVMs in your cloud, with the source code in your hands.
| engrams | Devin (Cognition) | Factory (Droids) | |
|---|---|---|---|
| Who runs it | You, in your own cloud account or on your own hardware | Cognition; on Enterprise, a single-tenant VPC that Cognition manages | Factory; on-premise and air-gapped options on Enterprise contracts |
| Source code | Open source. Read it, change it, and run every line | Proprietary | Proprietary |
| The agent | Claude Code, Codex, or your own harness, mounted by the host at boot; never baked into your image | Devin, Cognition's agent | Droids, Factory's agent |
| Models | Whatever your harness calls, on your own provider keys or model router | From Cognition's model list | From Factory's model list |
| Where the work runs | A Firecracker microVM per session, on your KVM nodes | A vendor-managed cloud machine per session | Vendor-managed cloud machines, or your machine through the CLI |
| Your environment | Any Linux image with /bin/sh, built with docker build; warmed once and snapshotted | Set up inside the vendor's machine configuration | Set up inside the vendor's environment configuration |
| Credentials | Brokered: the egress proxy inserts each key outside the VM, and the agent never sees it | Stored on the vendor's platform | Stored on the vendor's platform |
| Automations | Workflows you edit as a graph, started by a schedule, a pull request, Slack, Linear, or a webhook | Tasks from Slack, Linear, GitHub, and the API | Tasks from Slack, Linear, the IDE, and the CLI |
| Where people work | Dashboard, Slack threads, GitHub pull request review, CLI, API | Web app, desktop app, CLI, Slack, Linear | CLI, IDE, web app, Slack, Linear |
| What you pay | No license fee. You pay your cloud provider and your model provider | Per-seat plans with usage quotas | Per-seat plans with usage limits |
Vendor details are from their public product and pricing pages as of October 2026. Plans change; check with each vendor.
Where else teams look.
"Self-hosted" means different things. For most products below, the vendor keeps the control plane: the agent loop, the session record, and often the transcripts. Only the execution runs on your machines. engrams runs all of it in your account: the orchestration, the database, the VMs, and the snapshots.
| Who runs what | Where the work runs | Pick it when | |
|---|---|---|---|
| OpenHands | Open source agent and runtime; the hosted cloud, or self-hosted on Teams and Enterprise plans | Docker containers | You want the OpenHands agent itself, and container isolation is enough |
| Claude Code self-hosted environments | Anthropic runs the control plane and stores transcripts; runners on your hosts do the work (Team and Enterprise, beta) | Your hosts; isolation is yours to set up | You use only Claude and accept a control plane at Anthropic |
| Cursor self-hosted cloud agents | Cursor runs the agent loop and inference; your workers run commands and edits (team pools on Enterprise) | Your machines or Kubernetes pods | Your team works in Cursor and accepts a control plane at Cursor |
| OpenAI Codex cloud | OpenAI runs everything | OpenAI-hosted environments | You pay for ChatGPT, and hosted execution is acceptable |
| E2B | A sandbox SDK, not an agent product; open source infrastructure, or E2B runs it in your account | Firecracker microVMs | You are building your own agent product and need only the sandbox layer |
| Modal Sandboxes | Modal runs everything | gVisor containers or VMs, with GPUs | You want serverless sandboxes with GPUs and no infrastructure to run |
From each vendor's public docs and pricing pages as of October 2026. Plans change; check with each vendor.
Pick the other thing when it fits.
If you do not want to run infrastructure, a hosted product is the right call. Devin and Factory give you an agent in minutes and handle the machines and billing. engrams asks you for a Kubernetes cluster with a KVM node pool, a Postgres, a bucket, and someone to run them.
If you want one vendor's agent and models, tuned together and supported by that vendor, buy that vendor's product. engrams runs the agents you bring and does not ship a model of its own.
If your hosts cannot run KVM, engrams has no production mode for you. It does not run sessions as containers.