Coder: Self-Hosted Developer Workspaces
The self-hosted CDE platform where you run the control plane, the database, and every workspace on infrastructure you already own
Editions, license, deployment model, and cost - checked against Coder's own documentation as of September 2026
What It Is
Coder is a platform for provisioning and managing remote development environments on infrastructure you control. It is built and sold by Coder Technologies Inc. The product has one defining structural choice: the server component, called coderd, is software you install and operate. There is no vendor-hosted control plane in the picture, and no version of the product where Coder runs your workspaces for you.
The second defining choice is that Coder does not invent its own provisioning language. A workspace is described as a Terraform template, and the platform runs Terraform to build it. That is why Coder is unusually indifferent to what a workspace actually is. Coder's documentation names Docker containers, Kubernetes pods, AWS EC2 instances, Azure, and Google Cloud as provisioning targets. Because the layer doing the work is Terraform rather than a Coder-specific runtime, the practical target list is whatever your team can express in a Terraform provider.
The cost of that flexibility is that Coder hands you an infrastructure product, not a finished service. Somebody on your side owns the server, its PostgreSQL database, the ingress, the upgrade path, and the Terraform templates that define what a workspace looks like. Teams that want a development environment to appear without an infrastructure owner behind it are not the audience for this tool.
Who It Suits, and Who It Does Not
A Good Fit
- Organizations with a real platform or infrastructure team that already operates Kubernetes, Terraform, and PostgreSQL in production.
- Regulated environments where source code and build artifacts must never leave your network or your cloud account.
- Teams whose workspaces are not uniform - a mix of containers for web work and full VMs for kernel, embedded, GPU, or licensed-toolchain work.
- Air-gapped or on-premises deployments where a SaaS control plane is disqualifying on its own.
- Buyers who want the exit option that an open-source, self-operated control plane provides.
A Poor Fit
- Small teams with no infrastructure owner. You are adopting a server, a database, and an upgrade cadence, and somebody has to hold them.
- Anyone who wants a hosted service. There is no Coder-operated SaaS tier to fall back on.
- Teams with no Terraform skills and no appetite to acquire them. Templates are the product surface, and you will be editing them.
- Buyers who need a published price to get through procurement. Premium pricing is quote-only.
- Organizations needing multi-organization access control, audit logging, or high availability on a zero budget - those are Premium features, not Community ones.
Deployment Model
"Self-hosted" is used loosely across this market, so it is worth being exact. With Coder, both halves are yours. You operate the control plane, and you operate the compute the workspaces run on. Coder Technologies does not run any part of the serving path.
That distinction is what separates Coder from most of the products it gets compared against. A vendor that hosts the control plane and schedules workspaces into your cloud account is offering something real, but it is not the same thing, and it has different failure and compliance properties. With Coder, if the vendor's website went dark tomorrow, your installation keeps running until you choose to change it.
| Component | Who Operates It | Notes |
|---|---|---|
| Control plane (coderd) | You | API, dashboard, reverse proxy for workspace URLs, agent registration |
| PostgreSQL database | You | Holds all platform state - back it up like any other production database |
| Provisioner (provisionerd) | You | Separate execution context that actually runs Terraform to build and destroy workspaces |
| Workspace compute | You | Your cloud account, your cluster, or your own hardware - billed by your provider, not by Coder |
| Workspace agent | You | Runs inside each workspace, providing SSH, port forwarding, and health checks |
| Anything hosted by Coder | Nothing in the serving path | Support, licensing, and the module registry are vendor-side; your workspaces are not |
License and Open Source Reality
The coder/coder repository carries the GNU Affero General Public License v3, confirmed in the repository's own LICENSE file as of September 2026. This is the whole platform, not a stripped demo build: the Community edition gives you unlimited workspaces, unlimited templates, and unlimited members within a single organization.
So the honest answer on production viability is yes, with named gaps. Community is a real, usable platform. What it does not include is the enterprise operational layer - multi-organization access control, audit logging, quotas, high availability, workspace proxies, custom roles and branding, and a support SLA. Coder's pricing page also limits Community to a single platform integration and a single external authentication integration, which is a constraint worth checking against your identity and cloud footprint before you plan around the free edition.
AGPL-3.0 is a copyleft license with a network clause. For the overwhelming majority of adopters running Coder internally this changes nothing, but if you intend to modify Coder and offer it onward as a service, read the license terms properly rather than assuming permissive behavior. The Premium feature set is governed separately by a commercial license rather than by the AGPL grant.
Pricing
As of July 2026, coder.com/pricing lists exactly two editions. The older three-tier arrangement of OSS, Enterprise, and Premium is gone, so comparison articles that still describe three Coder tiers are out of date.
Coder does not publish a Premium price. It is described only as billed annually per user, with a contact-sales motion. Any specific per-seat figure you find elsewhere is somebody's estimate, and this page will not repeat it.
Community
Free (as of July 2026)
Open source and self-hosted. Unlimited workspaces, templates, and members within a single organization, plus SSO via OIDC. Limited to one platform integration and one external authentication integration. Support is community-based, through Discord and GitHub.
Premium
Not published
Billed annually per user, quote only, as of July 2026. Adds multi-organization access control, audit logging, quotas, high availability, workspace proxies, custom roles and branding, and an SLA. AI Governance - AI Gateway and Agent Firewall - is included with a Premium license rather than sold separately, and Coder Agents draws on a purchased pool of Agent Hours.
The license fee is not the whole cost
Because you run every workspace, your cloud bill is a first-class line item and it is the one that usually surprises people. Budget for the compute, the storage that persists between sessions, egress, and the engineering time to own the control plane. Auto-stop policies matter more here than the seat price does.
Architecture and Isolation
Coder's documented architecture separates three concerns. coderd is a deliberately thin API that connects users, workspaces, and provisioners, keeping its state in PostgreSQL and serving the dashboard, the HTTP API, and the reverse proxy for workspace URLs. provisionerd is a separate execution context that runs Terraform. The workspace agent runs inside each workspace and presents a consistent interface - SSH, port forwarding, health checks, startup scripts - no matter what operating system or cloud is underneath.
The consequence for isolation is important and often misread. Coder does not implement its own sandbox. The isolation boundary is whatever your Terraform template provisions. If the template creates a Kubernetes pod, you get pod-level isolation and its known caveats around shared kernels. If it creates an EC2 instance or a dedicated VM, you get hypervisor-level separation. If it targets your own hardware, you get whatever that hardware enforces. Coder gives you the choice; it does not make it for you, and it will not save you from a permissive template.
That separation of provisioner from control plane also has a practical benefit worth knowing about: provisioners can run in a different network context from the server, which is how teams reach into environments the control plane itself is not allowed to touch.
Status and Durability
Coder Technologies Inc. is a private, venture-backed company. Its about page names Ammar Bandukwala and Kyle Carberry as co-founders, Rob Whiteley as chief executive, and lists Uncork Capital, Founders Fund, IQT, Notable Capital, Redpoint, Capital Factory, and Georgian Partners as investors. The company also raised a 90 million dollar Series C led by KKR, announced 04-01-2026, with Qube Research and Technologies and Uncork Capital participating - a funding event worth weighing in any durability assessment, since it buys runway the smaller vendors in this category do not have. The customer logos and the "over 100,000 community users" figure on the about page are vendor claims, published by the company selling the product, and should be treated as marketing rather than independent measurement.
Maintenance activity is genuinely strong and easy to check for yourself. Coder cuts a release from main on the first Tuesday of each month and patches it continuously. As of September 2026 the published release calendar puts 2.37 on mainline, released 09-01-2026; 2.36 on stable, released 08-04-2026; 2.35 on security support; and 2.34 and 2.29 on Extended Support Release. Coder's own release notes advise enterprise customers without a staging environment to install the stable line rather than mainline, which is a useful signal about how to plan your upgrade policy. Patch numbers move within days, so plan against the release line rather than a specific tag.
The most important durability item has a date on it. Coder Tasks is deprecated and, in the shipped v2.37 line, hidden by default: the Tasks API at /api/v2/tasks is only registered when CODER_ENABLE_AI_TASKS is set, which is verifiable in v2.37.3's own source. The removal itself - the CLI verbs, the API and the RBAC resource - is merged on the main branch and ships in v2.38, scheduled for 10-06-2026, so a deployment that still calls the Tasks API has one release left rather than none. Coder's migration guide on main states that requests to the Tasks API fail and directs readers to a previous documentation version if they need to read about Tasks at all. Note what "removed" does and does not mean: it is removal from new releases, not from every deployment on the same day. Premium customers on the 2.34 Extended Support Release, which began 06-02-2026 and runs twelve months, keep Tasks until that line ends. Teams not on Premium have no such runway, and for them the migration is now overdue rather than upcoming. The documented replacement is Coder Agents, which entered beta on 05-06-2026 and reached general availability in that same v2.37 release. It runs entirely self-hosted with no SaaS component and works with any LLM provider you configure. The two are architecturally different rather than drop-in equivalents, and the difference is a security argument as much as an operational one: Tasks ran the agent inside the workspace with LLM credentials injected into the workspace environment, while Agents runs the agent loop in the control plane and, in Coder's words, never lets those credentials enter the workspace. The replacement API is Chats at /api/v2/chats, which keeps conversation state in the Coder database rather than in the workspace and supports multi-turn history and sub-agent delegation. If you are evaluating Coder for agent workloads, build on Agents.
Agent work is now separately metered, which is a cost line that did not exist when this category was only about workspaces. Coder's licensing documentation describes Community licenses as supporting up to five concurrently active agents deployment-wide, with no limit on how long they run and additional agents queued behind them, and Premium licenses as including a preset pool of Agent Hours shared across the deployment with unlimited concurrency. Agent Time is measured as the cumulative duration of model invocations that produce chat messages - it excludes time spent composing messages, waiting between turns, and failed or retried calls. Budget for it separately from seats, because it scales with how much autonomous work you run rather than with headcount.
Two AI features were also renamed: AI Bridge is now AI Gateway, and Agent Boundaries is now Agent Firewall. Coder describes the AI Gateway change as a cosmetic rebrand that nonetheless moves user-visible names, configuration options, the canonical HTTP API path, and the Prometheus metric names. Deprecated aliases still resolve for gateway and proxy controls - CODER_AIBRIDGE_*, --aibridge-*, the aibridge: YAML group, /api/v2/aibridge and the coder_aibridged_* metrics - with one exception that will bite an upgrade: provider configuration through environment variables, flags, and YAML has been removed under both the old and the new name, and is now managed in the database. New settings get only the ai_gateway name with no alias, so the old names are a migration window rather than a permanent equivalence. As of September 2026 the coder/aibridge and coder/boundary repositories both still exist under their original names and neither is archived. Agent Firewall's documentation carries the same warning that some configuration options and internal references still use the old name. Both features are part of AI Governance, which is included with a Premium license rather than sold as a separate add-on; Community deployments cannot access either.
On the exit question, Coder is in a better position than most of this category. Because the platform is AGPL-3.0 and you already operate every component, a change of vendor direction is a support and roadmap problem rather than a service shutdown. Your templates are Terraform, which you would keep regardless. That is a materially different risk profile from a hosted CDE, where a strategy change can take your environments with it.
Related Reading
Other vendors in this directory, and the wider selection question
Coder vs GitHub Codespaces
The control plane you operate against the one GitHub operates, on isolation, pricing, and exit
Coder vs Ona
Two different answers to where the orchestration plane lives, and what bring-your-own-cloud really buys
Coder vs Okteto
A workspace platform against a Kubernetes inner-loop tool - they answer different questions
Self-Hosted vs Managed
The decision this page sits inside, before you compare any two products
Okteto
The other self-hostable option, where only the CLI is open source and the platform is commercial
Eclipse Che
Foundation-governed rather than single-vendor, if that is the property you are actually buying
DevPod
A client rather than a platform - the comparison to make if you do not want a control plane at all
CDE Tools Comparison
Every platform side by side, including the commercial and discontinued ones
Build vs Buy
Whether you have the operational capacity that running a control plane really requires
Cost Analysis
Modeling the compute bill that a self-hosted deployment puts on your side of the ledger
Open Source CDEs
What "open source" actually covers in each product, license by license
Vendor Evaluation
A structured way to run the comparison, including the questions vendors dislike
Sources
Every figure on this site, with its publisher, date, and link
