Agentic Data Plane

rpk ai run opencode

Launch OpenCode 2 with model traffic routed through the Redpanda AI gateway for the current environment.

rpk ai layers native v2 providers, model settings and policies over your OpenCode configuration. A private standalone server owns this session; existing background servers are not used or changed. Conflicting selected-provider configuration, credentials and provider-use policies are refused before API-key acquisition. Other agents, tools, MCP servers and themes remain available. User plugins are trusted code, not an egress security boundary.

Each model uses the first supported wire API the server catalog declares:

openai_responses         direct           @opencode/ai/providers/openai/responses
                                            {provider}/responses
anthropic_messages       direct           @opencode/ai/providers/anthropic
                                            {provider}/v1/messages
gemini_generate_content  direct           @opencode/ai/providers/google
                                            {provider}/v1beta/models/{model}:streamGenerateContent
openai_chat_completions  direct           @opencode/ai/providers/openai-compatible
                                            {provider}/chat/completions
gemini_generate_content  vertex           @opencode/ai/providers/google
                                            {provider}/v1beta/models/{model}:streamGenerateContent
anthropic_messages       vertex           @opencode/ai/providers/google-vertex/messages
                                            {provider}/v1/projects/{p}/locations/{l}/publishers/anthropic/models/{model}:streamRawPredict
bedrock_converse         bedrock_runtime  @opencode/ai/providers/amazon-bedrock
                                            {provider}/model/{model}/converse-stream
openai_responses         bedrock_mantle   @opencode/ai/providers/openai/responses
                                            {provider}/openai/v1/responses
openai_responses         bedrock_runtime  @opencode/ai/providers/openai/responses
                                            {provider}/openai/v1/responses
openai_chat_completions  bedrock_mantle   @opencode/ai/providers/openai-compatible
                                            {provider}/openai/v1/chat/completions
openai_chat_completions  bedrock_runtime  @opencode/ai/providers/openai-compatible
                                            {provider}/openai/v1/chat/completions

A model needs streaming, tools, multi-turn conversations and system prompts. Unknown models and unsupported bindings are omitted; naming one with -m fails. OpenAI-compatible providers require an explicit model allowlist and use Chat Completions. Fireworks also requires an explicit allowlist.

Every bound model stays available through /models. -m optionally selects the starting model; otherwise the first bound allowlisted or catalog model is used. Title and compaction agents use that same initial model. A private fixed model catalog and request-time guard prevent other provider/model selections.

The gateway signs Bedrock and Vertex requests upstream; no local AWS or Google credentials are needed. The Agentic Data Plane API key is supplied only through the child environment, never in argv or literal generated configuration. OpenCode expands its environment placeholders, so its resolved local configuration may contain credentials. Do not publish that output. --print-config prints only the generated placeholder configuration without acquiring a key or starting OpenCode.

ChatGPT subscriptions: use an OpenAI passthrough provider whose upstream is https://chatgpt.com/backend-api/codex. Sign in with OpenCode using opencode auth login openai --method chatgpt-browser or --method chatgpt-headless. Token-sharing logins are not supported for this Codex backend. The native openai provider owns seat-token refresh and account headers; the Agentic Data Plane key travels separately in X-Redpanda-Api-Key.

Claude subscriptions: an Anthropic passthrough provider uses OpenCode native claude-agent and your native claude subscription login (claude 2.1.283 or newer). No external adapter is installed and no login is copied. OpenCode bridges its own tools to Claude; its permission controls and --auto apply. Subscription use is subject to your provider terms and seat limits.

Requires OpenCode 2.2.7 or newer (v2 only). Server/attach overrides and commands that bypass the standalone model guard are refused. Start from the intended project directory; forwarded directory, worktree and resume selectors are refused. Pass OpenCode arguments after a literal --:

rpk ai run opencode -L openai -- run "explain this repository"

Usage

rpk ai run opencode [-- OPENCODE_ARGS...] [flags]

Flags

Value Type Description

--api-key

string

authenticate the session with an Agentic Data Plane API key: ask (ask before creating one), yes (create one without asking, also without a terminal), no (do not create a key; require an already provided Agentic Data Plane API key). Default: ask.

-L, --llmprovider

name

AI gateway LLM provider name to route through.

-m, --model

string

Starting model ID. It must be in any configured allowlist and declare a supported wire API (default: the first bound model). Every bound model stays available through /models.

--print-config

bool

Print the generated OpenCode configuration, then exit without launching OpenCode or creating an API key. Provider read access is required.

Global flags

Value Type Description

--config

string

Redpanda or rpk config file; default search paths are ~/.config/rpk/rpk.yaml, $PWD/redpanda.yaml, and /etc/redpanda/redpanda.yaml.

-X, --config-opt

stringArray

Override rpk configuration settings; -X help for detail or -X list for terser detail.

--ignore-profile

bool

Ignore rpk.yaml and redpanda.yaml; use default settings.

--profile

string

rpk profile to use.

-v, --verbose

bool

Enable verbose logging.