You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
While Coder Agents is a nice built-in harness and provides many benefits (centralized control, auth, integrations, APIs), many users would like to see improved support for running other harnesses such as Claude Code, Cursor, Codex, OpenCode, or even harness orchestration tools like Onmiagent or Vibekanban on Coder Workspaces.
This can be to several reasons:
Cost: Harnesses like Claude Code offers per-user subscription pricing that makes AI much more affordable both for solo users and at scale versus token pricing which ends up being more expensive and less predictable.
Capabilities: The harness itself has more features (e.g. memory, sandboxing, desktop clients, etc) that Coder Agents does not have, but are important to user's workflows
Intelligence: The harness seems to "feel" more intelligent/capable of solving the task (even compared to Coder Agents with the same model) due to advanced harness development techniques, agent delegation, etc.
Cost of Switching: Some teams have built entire ecosystems around a specific harness, have strong adoption and do not see a clean migration path into Coder Agents
Current State of External Harness Support
We currently provide Agent Modules in the registry for the CLI agents which can install, authenticate, and configure an agent directly in a workspace.
This is valuable and can provide users a "golden path" into a workspace with a given harness, but it does not allow for "orchestration of agentic tasks" on workspaces. For that, we offer a REST API and MCP Server for "DIY" solutions but provide no direct UI or API for interacting with agentic sessions. Similarly, the workspaces themselves (via UI and API) are unaware of agents within workspaces and the auto-start and auto-stop mechanics of the workspace do not factor ongoing tasks.
Ongoing Work Tracker
Modules for CLI and IDE Agents in Coder Registry: Shipped ✅
Support for Self-Hosted Runners for platforms like Cursor: ⌛ In Progress
Link workspaces to the corresponding harness (sessions): Discussed, Not Planned Yet
Improve workspace lifecycle (start/stop) to be aware of other agents/harness activity: Not Planned Yet
Looking for Input
We'd love to hear from you on this topic:
Are you currently using Coder Agents, or an external harness with Coder? Why?
Where is would you specifically like to see us invest in supporting external harnesses?
While Coder Agents is a nice built-in harness and provides many benefits (centralized control, auth, integrations, APIs), many users would like to see improved support for running other harnesses such as Claude Code, Cursor, Codex, OpenCode, or even harness orchestration tools like Onmiagent or Vibekanban on Coder Workspaces.
This can be to several reasons:
Cost: Harnesses like Claude Code offers per-user subscription pricing that makes AI much more affordable both for solo users and at scale versus token pricing which ends up being more expensive and less predictable.
Capabilities: The harness itself has more features (e.g. memory, sandboxing, desktop clients, etc) that Coder Agents does not have, but are important to user's workflows
Intelligence: The harness seems to "feel" more intelligent/capable of solving the task (even compared to Coder Agents with the same model) due to advanced harness development techniques, agent delegation, etc.
Cost of Switching: Some teams have built entire ecosystems around a specific harness, have strong adoption and do not see a clean migration path into Coder Agents
Current State of External Harness Support
We currently provide Agent Modules in the registry for the CLI agents which can install, authenticate, and configure an agent directly in a workspace.
This is valuable and can provide users a "golden path" into a workspace with a given harness, but it does not allow for "orchestration of agentic tasks" on workspaces. For that, we offer a REST API and MCP Server for "DIY" solutions but provide no direct UI or API for interacting with agentic sessions. Similarly, the workspaces themselves (via UI and API) are unaware of agents within workspaces and the auto-start and auto-stop mechanics of the workspace do not factor ongoing tasks.
Ongoing Work Tracker
Looking for Input
We'd love to hear from you on this topic: