Docs

Supported runtimes and deployment status

Compare execution, approvals, steering, attachments, and isolation across runtimes.

A runtime appearing in configuration does not guarantee that it can execute a thread. This page records the runners and capabilities implemented in Convoy. Your machine also needs the runtime installed and authenticated.

Runtime execution

RuntimeThread executionApprovals and plan modeMid-run steeringInline image inputConvoy --isolated
Claude CodeSupported through the Agent SDKInteractive approvals and plan modeSupportedJPEG, PNG, GIF, WebPSupported
CodexSupported, app-server by defaultInteractive approvals, questions, and plan mode on app-serverApp-server onlyFiles referenced by local pathSupported
OpenCodeSupported through opencode serve and its SDKInteractive approvals, questions, and plan modeNot supportedFiles referenced by local pathRejected
CursorSupported through cursor-agent print modeNon-interactive Act and Guarded modes; no approval prompts or plan modeNot supportedFiles referenced by local pathRejected
Amp, Droid, Gemini, CodeRabbit, GitHub Copilot, OtherNo registered CLI runnerUnavailableUnavailableUnavailableUnavailable

Steering sends a queued message into an active turn. Without steering, follow-up messages wait for the next turn. Inline image input means Convoy passes image content directly to the runtime. A path reference instead requires the runtime to read the downloaded file with its own tools.

--isolated requests an OS sandbox through Convoy. Runtime permission modes are separate controls. OpenCode and Cursor fail an isolated run instead of silently running without that sandbox. See Isolation options.

Codex transports

The default app-server transport requires codex-cli 0.142.0 or later. It supports approval prompts, questions, plan review, steering, live text deltas, and the context-window meter.

convoy connect --codex-exec selects codex exec --json. Convoy also falls back to this transport with a warning when it detects an older Codex binary. Exec supports execution, model and reasoning selection, and resume. It does not support mid-run approvals, questions, plan review, steering, live token deltas, or the context-window meter. Approval-style modes degrade to sandboxed execution that never asks; Plan mode does not become read-only on this transport.

OpenCode and Cursor

OpenCode requires version 1.14.19 or later. Its connection discovers the machine's model catalog and per-model variants. It supports streaming output and session resume.

Cursor runs a fresh print-mode process for each turn and resumes the saved chat for follow-ups. Act allows commands through --force; Guarded follows the machine's permission configuration and skips denied commands. Neither mode asks for approval mid-run. Cursor streams coalesced output but does not report a context-window meter.

See Adding a runtime for the steps needed to add a runner. These capabilities were checked against the CLI runtime registry and runner implementations; they are not a claim that every runtime and deployment combination has been tested end to end.

Product guides

Workflows has its own guide for authoring and running stages, lanes, and steps. Watch Tower's public docs scope needs verification before a guide is added.

Planned / experimental deployment paths

Docker deployment describes the self-hosted path and its validation status. The managed path starts with Choose a deployment path. Follow the validation notes for the recipe you choose.

How to interpret "planned" copy elsewhere

Use this page to check runtime execution and capability limits. Integration configuration, runtime execution, and deployment validation describe separate things; a working configuration file alone does not establish all three.

On this page