SDK selection
Choose the SDK backend for code generation.
Reverse API Engineer can drive code generation through Claude, OpenCode, Codex, Copilot, or Cursor.
Claude (default)
Direct integration with Anthropic's Claude API. Uses your Anthropic API key. Recommended for most users.
{
"sdk": "claude",
"claude_code_model": "claude-sonnet-4-6",
"claude_setting_sources": []
}Claude sessions don't load your Claude Code settings files, CLAUDE.md, hooks, plugins, permission rules or MCP servers, so your personal setup doesn't steer the agent. Organization-managed policy still applies, and Claude Code still reads its own state such as ~/.claude.json and auto memory. If ~/.claude/settings.json provides your authentication (apiKeyHelper, or Bedrock, Vertex or proxy variables under env), set claude_setting_sources to the layers to load, e.g. ["user"] or ["user", "project", "local"]. Each layer you list is loaded in full, including its hooks, plugins, permission rules and MCP servers.
OpenCode
Uses the OpenCode SDK running locally. Pick this if you want to route through your own model gateway, run open-weights models, or use a non-Anthropic provider.
{
"sdk": "opencode",
"opencode_auto_start": true,
"opencode_base_url": "http://127.0.0.1:4096",
"opencode_model": "big-pickle",
"opencode_npx_package": "opencode-ai@latest",
"opencode_provider": "opencode"
}RAE reuses an existing OpenCode server. If none is listening, it downloads and
starts the latest opencode-ai release through npx, then stops that child
process when RAE exits. A global OpenCode installation is not required.
Automatic setup needs Node.js 20+; the first run downloads the OpenCode npm
package and later runs use npm's cache. Fresh configurations use the free
opencode/big-pickle model by default.
Set opencode_auto_start to false when you manage the server yourself. The
environment variables OPENCODE_BASE_URL, RAE_OPENCODE_PACKAGE, and
RAE_OPENCODE_AUTO_START override the matching settings. Provider credentials
and model identifiers remain OpenCode configuration. Password-protected servers
use OPENCODE_SERVER_PASSWORD and optional OPENCODE_SERVER_USERNAME; these
variables are inherited by an auto-started server. See
OpenCode providers and
models.dev.
RAE validates the configured provider/model pair against OpenCode's connected provider catalog before creating a session. Invalid pairs fail immediately and include valid alternatives plus currently exposed free, tool-capable models. Existing compatible servers are reused; versions older than RAE's tested OpenCode release display an upgrade warning.
Ollama through OpenCode
In /settings, set the OpenCode provider to ollama. RAE connects to the
configured Ollama daemon (default http://127.0.0.1:11434) and offers the
installed models that advertise tool calling and at least 64k context. Picking
one updates both opencode_provider and opencode_model.
{
"sdk": "opencode",
"opencode_provider": "ollama",
"opencode_model": "qwen3.5",
"ollama_auto_start": true,
"ollama_base_url": "http://127.0.0.1:11434"
}RAE reuses a running Ollama daemon or starts ollama serve when the CLI is
installed. It passes an inline provider definition to an auto-started OpenCode
server, so it does not overwrite opencode.json. If you run OpenCode yourself,
that existing server must already expose the selected Ollama model; RAE reports
the exact remediation when it does not.
Model downloads are intentionally explicit because they can consume many
gigabytes: run ollama pull <model> yourself, then select it. Override daemon
discovery with RAE_OLLAMA_BASE_URL and disable startup with
RAE_OLLAMA_AUTO_START=0. See Ollama's
OpenCode integration for the
upstream context-window guidance.
Copilot
Uses the GitHub Copilot SDK when the optional Copilot dependency and token are configured.
{
"sdk": "copilot",
"copilot_model": "gpt-5"
}Cursor
Uses the Cursor SDK bridge. Cursor can also load broader Cursor setting layers so WebFetch/WebSearch and configured tools match your desktop setup.
{
"sdk": "cursor",
"cursor_model": "composer-2.5",
"cursor_web_search": true,
"cursor_setting_sources": null
}Codex
Uses the OpenAI Codex SDK with your Codex sign-in. Install the optional dependency, which brings its own Codex runtime (about 170 MB):
uv tool install "reverse-api-engineer[codex]"
# or
pip install "reverse-api-engineer[codex]"{
"sdk": "codex",
"codex_model": null,
"codex_effort": null,
"codex_home": null
}Sign in once with Codex, with codex login or the Codex app. RAE runs Codex
in its own home, ~/.reverse-api/codex, whose auth.json links to the one in
your Codex home (~/.codex, or $CODEX_HOME), so both share one sign-in.
Your Codex config.toml, MCP servers, hooks, skills and AGENTS.md are not loaded,
nor the apps connected to your ChatGPT account, such as Gmail or Drive,
so your personal setup doesn't steer the agent, and RAE's runs stay out of
your Codex history. Sharing works with Codex's default file credential store;
if your Codex keeps its sign-in in the system keyring
(cli_auth_credentials_store), set codex_home to your Codex home instead.
codex_model:nulluses Codex's default model for your account. In/settings, codex model lists the models your account offers.codex_effort:nulluses the model's default reasoning effort. In the session,/effortlists the levels the model offers.codex_home: a Codex home to run in as it is, such as"~/.codex", which loads all of it: its config, model providers, MCP servers, hooks andAGENTS.md. On Windows, RAE needs permission to create symbolic links to share your sign-in (Developer Mode); otherwise setcodex_hometo your Codex home.
Like RAE's other backends, the agent runs without Codex's sandbox or approval
prompts, so it can install dependencies, run the generated client, and start
agent-browser. agent --dry-run checks that the SDK is installed and that a
sign-in is there.
Switching SDKs
In the CLI:
/settings then "sdk"Or edit ~/.reverse-api/config.json directly.