Coder

Artificial intelligence writes a growing share of the world's software. Claude Code went from a research preview in February 2025 to writing 4% of all public GitHub commits by February 2026. Enterprises saw similar growth in adoption. In April 2026, 75% of all new code at Google was AI-generated and approved by engineers, up from 50% in fall 2025 and more than 25% in October 2024.

Source: SemiAnalysis

As the volume of AI-written code has grown, so has the autonomy of the tools producing it. Coding agents such as Claude Code and Codex, rather than merely suggesting code, can read files, write code, run shell commands, and call tools without stopping for approval. While this autonomy improves the speed at which developers work, it also means fewer opportunities for a human to catch an error prior to execution.

For example, in a controlled study of 1.1K paid professional testers published in August 2026, the testers caught a deliberately planted dangerous command only 13.6% of the time. Anthropic’s auto mode caught 89% of the same commands, which still lets roughly one in ten through. Examples of incidents that have resulted have included agents deleting remote git branches and uploading authentication tokens.

Enterprises have struggled to put AI projects into production. In a 2025 survey of AI projects, the share of companies abandoning most AI initiatives before production rose from 17% to 42% year over year, with cost, data privacy, and security among the top obstacles respondents cited. For coding agents, both cost and risk control depend on how many tokens each agent generates and which systems it can reach. For coding agents, the binding constraint on enterprise adoption is moving from how well an agent writes code to how completely an organization can control the cost and risk of agentic implementation.

Coder is a self-hosted platform that runs both human developers and AI coding agents inside workspaces on infrastructure the customer controls. Instead of relying on a developer to correctly approve individual actions, Coder provides controls like Agent Firewall to limit what agents can reach, keeps audit trails of token usage and actions tied to each developer, and applies the same controls to Claude Code, Codex, and its own agent. Coder's approach is to make spending, reach, and accountability properties of the environment agents run in, rather than properties of the agents themselves.

Kyle Carberry, Ammar Bandukwala (CTO), and John Andrew Entwistle (former CEO) founded Coder in Austin, Texas, in 2017. The founders met online while in high school and spent their free time building and monetizing Minecraft plug-ins and servers. They grew frustrated with how much time they spent moving code from their laptops to the game servers, where it had to run. In March 2019, the company released code-server, an open-source project that runs Microsoft’s Visual Studio Code (VS Code) on a remote server and opens it in a browser, so the editing, compiling, and running of code all happen on the machine that needs them. Coder then built its commercial enterprise product on top of that open-source project.

Entwistle served as CEO through Coder's early growth until January 2021. He and his co-founders then decided together that Bandukwala should take over as CEO. Entwistle went on to found Wander, a short-term rental startup, in May 2021 before leaving Coder's board in November 2021.

Bandukwala led Coder as CEO until May 2023, when he and Carberry brought on Rob Whiteley to fill the role. Whiteley had run F5's NGINX product group, where he built a commercial business on top of NGINX's open-source web server and worked to differentiate its free and paid tiers. Bandukwala and Carberry recruited Whiteley to accelerate growth after the open-source project and its commercial version had found initial, repeatable success.

In February 2024, Coder brought on Josh Epstein as chief revenue officer after he had grown HashiCorp’s global accounts business to more than $175 million. HashiCorp is the developer of Terraform, the infrastructure-as-code tool that Coder’s workspaces are built on. Coder promoted Epstein to president and chief business officer in October 2025, 20 months after he joined, in an announcement that cited 5x revenue growth. In September 2026, Whiteley stepped down as CEO and moved to Coder's board, and Epstein became CEO.

Source: Coder

As of October 2026, Coder’s product is built around running developers and AI agents in secure, self-hosted environments. At its core, Coder Workspaces act as a secure environment for developers and agents to work in. Coder Agents is Coder’s native AI coding agent, which runs in Coder’s control plane, the central server that manages workspaces and users, and creates Coder Workspaces on demand whenever it needs to execute a task, like writing code. Lastly, Coder AI Governance is a set of tools and controls that let customers audit and restrict any coding agent, whether Coder's own or a third-party agent.

Coder Workspaces

Coder’s Workspaces function as cloud development environments (CDEs) and are a direct expansion of the founders’ code-server project. In addition to the time it took developers to move code between personal machines and servers, the differences between developers’ devices caused crashes when running code. While developers may pull identical code from a central repository to run locally on their laptop, differences in operating systems and software versions can make an application crash on one laptop but run perfectly on another. Coder’s workspaces address this by letting an organization define a workspace once as a template and have every developer provision consistent copies from it.

Workspaces are fully configurable and can be tailored to specific project requirements. Those requirements are written into reusable templates in Terraform, a tool that lets an administrator describe the infrastructure they want in a configuration file rather than writing step-by-step code to create it. In a template, an administrator specifies the base infrastructure a workspace runs on (AWS, Kubernetes, and others) and which secrets it can use (API keys, Model Context Protocol servers, and others), so the person who writes the template controls how much freedom the workspace's users have. A developer can still connect through the code editor they work best with while the team keeps the uniformity it needs to work efficiently at scale.

Reproducible environments are not a new idea. Coder's version tracks each organization's templates and workspaces and automates their deployment, setup, and teardown through three components: coderd, provisionerd, and a workspace agent.

Coderd is a server backend that runs Coder's control plane. This control plane connects workspaces, provisioners, and users, and coderd is the only component that communicates with the PostgreSQL database, which holds session tokens, API keys, and the Terraform state for every workspace. Anytime a user starts up a workspace, coderd receives the request. Coder also schedules automatic startup and shutdown so a workspace shuts down when idle to save costs and spins up on a schedule, so it's already running by the time a developer needs it.

Provisionerd is a daemon, or a continually running background process, that runs Terraform during workspace builds.

Once a workspace is created, a workspace agent begins running inside it and gives coderd and connecting clients a consistent interface to the workspace. When a developer connects, the client and the agent establish an encrypted tunnel using WireGuard, an open-source VPN protocol, directly between the two where the network allows it, so commands and data sent from the developer's laptop reach the workspace securely.

Splitting the system into three processes also isolates risk. Coderd is the only process that touches the database, and provisioners can run builds in isolated environments that keep cloud secrets away from the control plane and stop a malicious template from reaching the Coder host. Additionally, since these are three independent processes rather than one program, each can be scaled without touching the others. Coderd scales for how much dashboard and API traffic it needs to serve. Provisionerd scales for build throughput, as an admin can run more of them to let more developers start workspaces simultaneously, independent of how busy coderd is. Lastly, because every workspace agent only manages its own single workspace, agents scale automatically as workspaces are created, with no coordination needed between them.

Coder Agents

Source: Coder

Coder launched Coder Agents in beta in May 2026 and made it generally available in September 2026. It is an AI coding agent whose reasoning loop runs inside the control plane rather than inside any individual workspace. This keeps chat history in the control plane when a workspace is stopped, and it keeps LLM credentials central, since the workspace holds no model API keys. Coder Agents is model agnostic. It is a standalone agent rather than a wrapper around third-party tools like Claude Code or Codex, and it connects to model providers including Anthropic, OpenAI, Google, and AWS Bedrock. Coder Agents only uses compute when a task needs it. The agent reads the organization's templates, picks one, and creates a workspace for the task.

Third-party agents such as Claude Code and Codex run inside a developer's Coder workspace instead. In September 2026, Coder launched Agent Relay in private preview, which lets cloud-hosted agents from Cursor and Claude Code execute their tool calls inside self-hosted Coder workspaces while the provider keeps the agent loop.

In the control plane, every action is tied to an authenticated identity through the customer's existing single sign-on stack. Platform teams can centrally approve which providers, models, and API keys are usable, set per-user and per-group usage limits and cost tracking, and restrict which templates agents are allowed to provision workspaces from. Coder also measures token consumption, model selection, pull-request activity, and merge rates across the organization. Coder Agents also integrates with GitHub Actions, Slack, and Jira, so agents can be triggered from outside the Coder interface.

Coder AI Governance

Source: Coder

In December 2025, Coder launched AI Bridge and Agent Boundaries, the two products behind AI Governance. It brought both to general availability in February 2026 as a paid add-on for Coder Premium and renamed them AI Gateway and Agent Firewall in July 2026. As of October 2026, AI Governance is included with a Premium license.

AI Gateway is hosted in the control plane and sits between any AI tool and the upstream model providers. Each LLM request is authenticated and tied to a named user, and prompts are centrally logged along with token usage and tool invocations.

Source: Coder

Agent Firewall is a process-level firewall enforced inside each workspace, restricting whatever AI tool is running there to an explicit allow list of domains and HTTP methods. Every allow or deny decision is logged to the control plane, where it can be exported to a customer's security information and event management (SIEM) system. Both Agent Firewall and AI Gateway integrate with a customer's identity provider through OpenID Connect (OIDC) and System for Cross-domain Identity Management (SCIM), allowing Coder to tie any given AI action back to a specific authenticated developer.

Customer

Coder's target customers are large, security-conscious companies where regulatory requirements prevent source code from leaving the organization's cloud or data center. Coder's self-hosted design keeps source code and data on infrastructure the customer already owns, which lets security teams treat it as an internal deployment rather than a third-party service handling the customer's code. Many of these companies rely on legacy systems that have become expensive or hard to standardize. A virtual desktop infrastructure (VDI), for instance, serves a similar function to Coder’s workspaces by giving each user a configured, centralized desktop operating system. However, a VDI often assigns each developer a dedicated virtual desktop, and those machines typically run 20-24 hours a day, on top of Windows licensing costs.

J.B. Hunt, a freight and logistics company, hosted its remote developers on more than 150 Azure virtual desktop instances before adopting Coder. This approach produced laggy remote sessions, up to two weeks of onboarding time per new hire, and inconsistent setups. After migrating, it cut developer VDI cloud spend by more than 90%, consolidating from over 150 individual VDI environments to a handful running on Kubernetes. Similarly, payabl., a payments company certified to the highest level of the payment card industry's data security standard, cut onboarding time from one week to under one minute and improved resource efficiency by 20% by automatically shutting down idle workspaces. Palantir used Coder to reduce environment configuration drift and compile times.

Market Size

The cloud development environment (CDE) market category is relatively new. In September 2024, CDEs had reached less than 5% of their potential target audience, and 30% of organizations in regulated industries are projected to mandate CDE use for security and compliance by 2028, up from 10% in 2024. In May 2026, a new market guide defined agentic development sandbox platforms as an emerging category.

The budget Coder competes for is moving toward AI coding. Enterprises spent $4 billion on AI coding tools in 2025, up from $550 million in 2024, and 90% of enterprise software engineers are projected to use AI code assistants by 2028, up from fewer than 14% in early 2024. One estimate put the VDI market, which includes the developer desktops Coder replaces, at $19.3 billion in 2025.

The rise in developer headcount drives demand for both AI coding tools and the environments they run in. As of Q3 2025, there were 48.4 million developers worldwide, with headcount jumping 21% from 2023 to 2024. Each developer running several agents multiplies the environments a platform team has to template and secure.

Competitive Landscape

Coder’s competition can be placed into two categories. The first is the cloud development environment, where it faces browser-based and hosted environments such as StackBlitz and GitHub Codespaces. Most of these run on the vendor's infrastructure, while Coder runs inside the customer's environment.

The second bucket of Coder’s competitors is a newer layer where AI agents execute code, where venture-backed sandbox providers like Daytona and E2B sell isolated environments for agents through an API, and where model providers have started building or buying the same capability. In 2026, OpenAI acquired Ona, formerly Gitpod, and Anthropic launched self-hosted sandboxes for its managed agents, with Daytona among the providers.

Coder's differentiation from its competitors is that it governs both humans and agents in one self-hosted system and is not tied to a single model provider. Its workspaces serve human developers as well as agents, and AI Gateway and Agent Firewall apply the same controls to Claude Code, Codex, or Coder's own agent. Sandbox providers sell execution for agents alone, and model providers' self-hosted options govern only their own agents.

Competitors

Cloud Development Environments

StackBlitz (Bolt.new): Founded in 2017 in San Francisco, StackBlitz built a browser-based development environment on its own WebContainer technology, which runs Node.js projects inside the browser. As of October 2026, it had raised $113.4 million in total from investors including Greylock, GV, and Madrona. It raised a $105.5 million Series B in January 2025, led by Emergence Capital and GV, at a $700 million post-money valuation. Its full-stack web development agent, Bolt.new, reached $40 million in ARR by March 2025, about five months after launch, after StackBlitz had nearly shut down. In September 2026, the company made its first acquisition, Dokai, to build agents for enterprise customers. StackBlitz avoids provisioning cloud infrastructure by running code client-side, while Coder is built to give enterprises control over that infrastructure.

Ona (formerly Gitpod): Gitpod was founded in 2020. It built an open-source platform for ready-to-code cloud workspaces and, prior to its acquisition, had raised $41 million across four rounds. Gitpod Classic shut down its usage-billed plan in October 2025 after the company repositioned around AI agents as Ona. OpenAI agreed to acquire Ona in June 2026 and closed the deal in August 2026, on terms that were not disclosed. The acquisition gives Codex a way to run agents in customer-controlled environments, which puts a former workspace rival inside one of the model providers whose agents Coder governs.

DevZero: Founded in 2021 in Seattle by former Uber engineers Debosmit Ray and Rob Fletcher, DevZero began selling production-like development environments for teams building on Kubernetes but then pivoted toward a Kubernetes optimization platform that reduces wasted cloud resources. As of October 2026, it had raised $26 million, split between a seed round led by Foundation Capital and a Series A led by Anthos Capital, with no new round since January 2023. As of October 2026, DevZero competes with Coder much less directly, since it sells infrastructure cost savings rather than developer workspaces.

GitHub Codespaces: Launched as a feature of GitHub, a Microsoft subsidiary, Codespaces entered beta in May 2020 and reached general availability for Team and Enterprise plans in August 2021. Codespaces provides cloud-hosted coding environments that open in a browser or VS Code and are created from a GitHub repository, so a developer can start coding without configuring their personal laptop. Microsoft's market cap was $3.9 trillion as of October 2026. Codespaces holds a distribution advantage due to its direct connection to GitHub, and in April 2026, GitHub added data residency in Australia, the EU, Japan, and the US for Codespaces on GitHub Enterprise Cloud, which addresses part of the regulated buyer's objection. It remains a proprietary managed service on GitHub's infrastructure, whereas Coder runs inside the customer's own environment.

Agent Sandboxes

Daytona: Founded in 2023 in New York, Daytona builds sandboxes for AI agents that it says start in under 90 milliseconds. In October 2026, Daytona raised a $60 million Series B at a $1 billion valuation, bringing its total funding to $91 million. In May 2026, Anthropic listed Daytona as one of four supported providers for self-hosted sandboxes in its managed agents product, then in public beta. Daytona sells sandboxes through an API and does not offer the human developer workspace or the identity-tied audit layer Coder bundles with them.

E2B: Founded in 2023 in San Francisco, E2B sells open-source sandboxes that AI agents use to run the code they generate, with an option to self-host inside a customer's AWS or GCP account. As of October 2026, it had raised $35 million, including a $21 million Series A in July 2025 led by Insight Partners, at an undisclosed valuation. E2B sells execution for agents through an API, without the human developer workspace or the governance layer Coder bundles with it.

Coder doesn’t provide the compute, storage, and infrastructure that power its customers' workspaces. Customers provision workspaces on infrastructure they already own, so those costs sit with the customer and Coder's own business is asset-light.

Source: Coder

Coder’s Community plan is directed toward hobbyists and small teams. It’s free and provides full access to Coder’s open-source core features, including unlimited workspaces, unlimited template creation, access to Coder Agents, and third-party agent compatibility. Premium adds prebuilt workspaces and exclusive access to AI Governance features such as AI Gateway, Agent Firewall, and user auditing, as well as cost control measures like usage quotas.

While Community users can create and store unlimited templates, spinning up a workspace off of a template’s specifications takes time as infrastructure must be provisioned, dependencies installed, and repos cloned. This process can take up to 15 to 20 minutes in complex cases. Premium, however, enables developers to keep a warm pool of workspaces that start in under a minute, reducing onboarding time and allowing developers to spawn short-lived workspaces for quick tasks.

Source: Coder

As of October 2026, Coder's pricing page caps both Community and Premium at five concurrently active Coder Agents across a deployment, while AI Premium removes the cap and adds agents triggered through an API and agent jobs that run in network-isolated workspaces. Five concurrent agents across an entire deployment is a hard ceiling for any team larger than a handful of developers. Coder's licensing documentation describes one way Premium buyers can exceed it. A Premium license includes a pool of "Agent Time" that lets agents run concurrently until the pool is used up, after which the five-agent cap applies again.

When Whiteley joined in 2023, Coder had just passed $1 million in revenue across 30 customers and had about 20 employees. Coder’s reported bookings grew 300% in the four quarters leading up to its April 2026 Series C, with 184% trailing 12-month net dollar retention, meaning existing customers' spending nearly doubled over the year. In September 2026, Epstein reported that revenue grew 400% in 2025, that net dollar retention was over 200%, and that Coder had more than 200 employees and hundreds of customers.

As of October 2026, Coder Premium is available on AWS Marketplace as part of an ISV Accelerate partnership, and Coder licenses are also sold through Google Cloud Marketplace. Coder signed a strategic collaboration agreement with AWS in July 2026. It also sells to public-sector buyers through Carahsoft, a government IT distributor.

Notable customers include Dropbox, KKR, Palantir, Mercedes-Benz, Square, J.B. Hunt, PsiQuantum, and Discord. Dropbox moved 1K developers onto Coder in four months and cut its cloud compute and storage costs by 30%. Coder holds a SOC 2 Type II report and has provided more than 2.5K developers at a US defense intelligence organization with standardized environments.

Coder also continues to maintain a substantial open-source footprint. Its code-server repository, which brings a browser-based version of VS Code to any machine, has 79.6K stars on GitHub as of October 2026, while its core coder/coder repository has 17K stars. Coder's GitHub organization has more than 240 public repositories as of October 2026.

In April 2026, Coder raised a $90 million Series C led by KKR, with participation from Qube Research & Technologies (QRT) and existing investor Uncork Capital. Both KKR and QRT are Coder customers, with KKR deploying Coder across more than 500 engineers and QRT running it for half of its more than 2K employees. Coder did not disclose a valuation for the round. One unverified estimate put its post-money valuation after the April 2026 round at $828 million. As of October 2026, Coder had raised a total of $172.8 million in funding.

The Series C followed a $35 million Series B extension round in June 2024, led by Georgian, with participation from existing investors Uncork Capital, Notable Capital, and Redpoint Ventures, at an undisclosed valuation. Before that, Coder raised a $30 million Series B in April 2020, led by GGV Capital, with participation from In-Q-Tel, a non-profit investment firm the CIA set up in 1999 to bring commercial technology to US national security agencies. That April 2020 round reportedly valued Coder at about $200 million.

Prior to this, Coder raised an $8.4 million Series A led by Redpoint Ventures in 2019, and a $4.5 million seed round in November 2018, co-led by Uncork Capital and Redpoint Ventures, with Founders Fund and Capital Factory participating. More than four years passed between the Series B and the subsequent round, which Coder raised in June 2024 as an extension that brought its Series B total to $65 million, and the Series C followed less than two years later as bookings accelerated.

Sovereign Infrastructure Mandates

Governments and regulators are pushing enterprises to keep software infrastructure under local control. More than 75% of European and Middle Eastern enterprises are projected to move virtual workloads to local options such as sovereign clouds, regional providers, or their own data centers by 2030 to reduce geopolitical risk, up from less than 5% in 2025.

Effective January 2025, the EU's Digital Operational Resilience Act requires banks, insurers, and other financial entities to manage the risk of their third-party information and communication technology providers. In June 2025, Microsoft announced a Sovereign Private Cloud designed to run in hybrid or air-gapped environments. As compute moves into these regional and isolated environments, the AI coding agents that run on top of it have to be deployable there too, which a tool offered only as a vendor-hosted service cannot do. Coder already supports air-gapped and offline deployments, and it plans to use part of its $90 million Series C to expand across Europe, Asia, and North America.

Governing Agents Beyond Engineering

Identity systems built for human employees and static service accounts are struggling to keep pace with AI agents. A February 2026 industry survey found that more than 16% of organizations do not track the creation of new AI-related identities and that 24% take more than 24 hours to rotate or revoke a credential after potential exposure.

Coder's customers are already extending it past engineering. In April 2026, KKR said it was extending Coder to analysts, data scientists, and other non-engineering users, and QRT had rolled Coder out to half of its more than 2K employees, a mix of engineers, data scientists, and analysts. As of October 2026, Coder's agents page frames AI use as already spanning marketing managers, designers, product managers, financial analysts, and HR leaders.

Coder's architecture is suited to this shift because it avoids creating a new credential type to govern. The agent loop runs in the control plane, LLM credentials never enter the workspace, and every agent action is tied to the authenticated user who submitted the prompt. The agent has the same permissions as that user, with no shared bot identity or service account. An analyst's agent has analyst-level access, and an HR manager’s agent has HR-level access. Coder also integrates with identity providers through OIDC and SCIM and exports audit events to SIEM tools, so it plugs into a customer's existing identity provider.

Agent Vendors Running Inside Customer Infrastructure

Coder's main value-add is that it's capable of fully running inside customer infrastructure. In October 2025, GitHub enabled its Copilot coding agent to run on an organization's own self-hosted runners instead of GitHub's cloud sandbox, giving it access to internal resources, such as packages, that are not reachable from the public internet, and in April 2026, GitHub added organization-level controls to make those runners the default. GitHub does not need to win enterprises away from an existing tool the way Coder does; it only needs to extend governance into a product teams are already using, which erodes the self-hosted argument Coder has used to differentiate from hosted-only competitors like Codespaces.

Model providers are moving the same way. In August 2026, Anthropic launched self-hosted environments for Claude Code in public beta, which run sessions inside the customer's network, and OpenAI's acquisition of Ona brings customer-controlled environments to Codex. Coder's Agent Relay, launched in September 2026, runs tool calls from Cursor and Claude Code inside Coder workspaces. If enterprises standardize on one agent vendor and accept that vendor's own self-hosted option, Coder's case narrows to customers that want one control plane across several agents.

Agent Governance at the Identity Layer

Enterprises may solve agent governance at the identity layer instead of through their development environment provider. In February 2026, Hush Security, a Tel Aviv-based startup, announced general availability of its Unified Access Management Platform, which replaces long-lived secrets with scoped, just-in-time credentials tied to verified workload identities. For AI agents, Hush calculates the merged permissions of an agent and the user it acts on behalf of, the same delegated-identity idea Coder uses, applied to downstream resources instead of the workspace.

Larger incumbents have moved to provide similar solutions as well. In August 2026, Okta announced general availability of Agent SSO, which registers AI agents that support its Cross App Access standard as identities in its directory, at no additional cost within core SSO plans used by more than 20K customers. Okta sells discovery of unregistered agents, named human owners, access certifications, and agent deactivation separately as Okta for AI Agents.

The threat to Coder is about budget ownership. Identity teams already control the budget and vendor relationships that agent governance would draw from, and Coder's own product integrates with identity providers through OIDC and SCIM, so those vendors sit upstream of it. If enterprises decide that discovery, ownership, and access policy for agents belong in the identity layer, Coder's remaining role narrows to runtime containment.

Supply Chain Exposure Through Coder's Registry

In August 2026, an attacker used a compromised Cloudflare API key to redirect part of the traffic to Coder's module registry for 14 hours, serving tampered modules designed to steal cloud credentials. Coder said its codebase and Google Cloud infrastructure were not compromised. Self-hosted deployments that pulled modules from that registry were still exposed to a server Coder, not the customer, controlled. Coder's pitch to security-conscious buyers rests on keeping code and credentials inside the customer's perimeter, so a repeat incident in its own supply chain would undercut the case that sets it apart from hosted alternatives.

AI coding agents are moving into enterprise development quickly, with a KKR managing director stating in April 2026 that more than 80% of enterprise developers were using or planning to use them. Similarly, projections state that by 2028, 30% of organizations in regulated industries would mandate cloud development environments. Coder sells a self-hosted platform that provisions those workspaces, and the isolated environments coding agents run in, on the customer's own infrastructure, with controls over what each agent can access. It raised a $90 million Series C led by KKR in April 2026 and reported that bookings grew 300% in the year before the round. As GitHub, Anthropic, and OpenAI add self-hosted execution and governance to their own agents, the open question is whether enterprises keep paying for an agent-agnostic control plane once each agent vendor offers its own.