DEV Community

Hamza
Hamza

Posted on Originally published at tekmag.thsite.top

Coder Agent Relay — Self-Hosted Execution for Cloud AI Coding Agents

#coder #agentrelay #cursor #claudecode

Originally published at https://tekmag.thsite.top/coder-agent-relay-self-hosted-execution-for-cloud-ai-coding-agents/

Coder's Agent Relay is a self-hosted execution environment that lets enterprises run cloud-hosted AI coding agents like Cursor and Claude Code inside their own secure infrastructure instead of the vendor's cloud. Launched September 4, 2026, it splits where an agent thinks from where it acts, so the reasoning loop stays in the vendor's cloud while every code read, command run, and file write happens on machines the company already governs.

Agent Relay answers the question that stalls most enterprise AI agent rollouts: not whether an agent is good enough, but where it actually executes. By default, a cloud-hosted coding agent reads your codebase, runs commands, and writes files inside infrastructure the vendor controls. For a bank, a hospital, or a government agency, that is a non-starter because source code, credentials, and customer data cannot leave the perimeter. The broader shift here is the same multi-vendor model reality developers are already living with, now reaching into where those models execute.

I verified this launch directly against Coder's announcement post and cross-referenced it with the Claude Code integration follow-up. The details hold up: Cursor shipped as the first integration partner on launch day, Claude Code followed eleven days later on September 15, and the product sits in private preview with early-access design partners. Sources differ on the launch date by two days: the original announcement is timestamped September 4, 2026, while the Claude Code post later refers to September 2. I'm treating September 4 as the launch date because that is the published timestamp on the primary announcement.

Key Takeaways

  • Agent Relay keeps a cloud agent's reasoning in the vendor's cloud but moves its execution into self-hosted Coder workspaces.
  • Cursor was the launch partner on September 4, 2026; Claude Code was added on September 15, 2026.
  • Agent Firewall, RBAC, audit logging, and ephemeral workspaces apply automatically to every agent action.
  • The product is in early access now, aimed squarely at regulated industries with strict data-residency rules.

What Agent Relay Is and How the Architecture Works

Agent Relay is a broker layer between a cloud-hosted coding agent and the enterprise infrastructure where that agent does its work. The reasoning loop, which includes LLM inference and decision-making, keeps running in the agent provider's cloud. The execution environment, where the agent reads code, runs tools, and writes files, runs inside a self-hosted Coder workspace on the customer's own VPC, on-premises hardware, or air-gapped network.

The mechanism is straightforward. The agent provider lets an organization configure "compute pools" that decide where agent work executes, and Coder registers itself as one of those pools. When a developer starts a session, the provider sends a resource request, and Agent Relay launches a workspace using prebuilt templates for a warm start. A secure connection then opens between the provider's daemon inside the workspace and its external service. Because the work happens inside a Coder workspace, governance applies with no extra setup.

One boundary is deliberate. LLM inference calls still route through the provider's infrastructure rather than through Coder's AI Gateway, a direct result of the reasoning loop living in the vendor's cloud. Coder acknowledges this plainly in its own announcement. Teams that need model access under their own control are pointed to Coder Agents, the fully self-hosted alternative that runs both reasoning and execution on enterprise hardware. For a look at how production teams are already structuring agent workflows with engineering-grade rigor, see our earlier coverage.

Security, Governance, and Identity

The security story is where Agent Relay earns its enterprise positioning. Every workspace is sandboxed with process-level isolation through nsjail (Linux namespaces) and LandJail (Linux Landlock v4), and networking defaults to deny. Only explicitly allowed outbound connections, such as an internal Git host or artifact registry, get through. "Agent Firewall, RBAC, and audit logging apply automatically, so there is no separate security model to build for agent traffic."

Identity mapping closes the other half of the governance gap. A sandboxed workspace only matters if you can trace which human triggered each agent action. Agent Relay resolves every incoming request against the organization's existing identity provider, so commits, tool calls, and file access carry the same attribution as a human developer's. Coder reports that one global financial institution using its identity and audit controls cut compliance audit resolution time from months to minutes.

Workspaces are also ephemeral. They are destroyed the moment a session ends, which means no lingering credentials or data after the fact. For regulated industries, that combination of default-deny networking, identity attribution, and short-lived environments is what moves an agent from "security will never approve this" to "here's the audit trail."

Coder's own AI Maturity Model assessment of 100 engineering organizations found that 70% are running AI agents on infrastructure never designed to support them, and only 31% have reached organization-wide governance. The blocker is usually the same: nobody can say which human an agent's actions trace back to.

How It Compares to the Alternatives

Agent Relay occupies a specific niche that other products don't fully cover. Gartner projects 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5% in 2025, and most organizations are running several agents from different vendors rather than standardizing on one. Coder's open-source platform has long offered self-hosted developer workspaces, and competitors approach the AI agent problem from different angles.

GitHub Codespaces provides self-hosted dev environments, but it isn't built for routing an external agent's execution into a governed workspace. Amazon Q Developer and similar vendor-hosted tools keep everything in one vendor's cloud, which locks teams to a single provider. Cognition's Devin is fully cloud-hosted with no self-hosted option at all. Agent Relay's differentiator is that it works with whatever agent a team already chose, supporting multiple providers like Cursor and Claude Code inside the same governed execution model.

Coder's own Coder Agents product is the closest sibling and worth understanding as a contrast. Coder Agents runs both the reasoning loop and the execution on enterprise infrastructure. Agent Relay only relays execution. The choice comes down to governance tier: Agent Relay is for teams that have already picked their agent and need to make it governable, while Coder Agents is for teams that need the full stack self-hosted, including air-gapped environments.

There are also lighter-weight CLI tools like Continue.dev and Aider, but those handle a single developer's local workflow rather than enterprise-scale governance, auditing, and multi-tenant workspace management. The gap Agent Relay fills is real, and no one else is attacking it head-on. It is part of a larger pattern of AI agents moving into governed, observable infrastructure rather than running loose in the cloud.

Conclusion

Agent Relay changes the procurement math for AI coding agents. Instead of waiting for a single agent to win, or rebuilding governance for each new tool, it separates the two problems: the vendor keeps the agent, the enterprise keeps the execution. "Separating where an agent thinks from where it executes is what makes cloud-hosted agents acceptable to security and compliance teams in the first place."

What I'll watch next is whether the provider ecosystem broadens. Coder has said any self-hosted cloud agent can build onto Agent Relay over time, and the fast follow from Claude Code, just eleven days after launch, suggests that is already happening. The companies most eager for this, regulated industries with strict data-residency requirements, are also the ones with the most to gain and the least room to get it wrong.

What are your thoughts on running cloud coding agents inside self-hosted infrastructure? Share your take in the comments.

{
"@context": "https://schema.org",
"@type": "FAQPage",
"mainEntity": [
{
"@type": "Question",
"name": "What is Coder Agent Relay?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Agent Relay is a self-hosted execution environment from Coder that lets enterprises run cloud-hosted AI coding agents, such as Cursor and Claude Code, inside their own secure workspaces rather than the vendor's cloud."
}
}
]
}

Frequently asked questions

What is Coder Agent Relay?

Agent Relay is a self-hosted execution environment from Coder that lets enterprises run cloud-hosted AI coding agents, such as Cursor and Claude Code, inside their own secure workspaces rather than the vendor's cloud.

When did Agent Relay launch, and which agents does it support?

Agent Relay launched September 4, 2026 with Cursor as the first integration partner. Claude Code support was added on September 15, 2026, and Coder plans to extend the integration to more providers over time.

Is Agent Relay generally available?

No. Agent Relay is in early access and private preview with select design partners. Teams can contact Coder's sales team to request access, and the documentation is public for evaluation ahead of broader availability.

How is Agent Relay different from Coder Agents?

Coder Agents runs both the reasoning loop and execution on enterprise infrastructure. Agent Relay only relays execution, keeping the reasoning loop in the vendor's cloud. Agent Relay suits teams that have already chosen their agent; Coder Agents suits teams that need the full stack self-hosted.

Does Agent Relay handle LLM inference on my infrastructure?

No. LLM inference calls still route through the agent provider's infrastructure rather than Coder's AI Gateway. Only the execution environment, where code is read and commands run, sits on the customer's own infrastructure.


References

Top comments (0)