OpenCode setup
Use Batchwell in OpenCode by adding a custom provider to ~/.config/opencode/opencode.json with settings.baseURL set to the Batchwell base URL.
To use Batchwell in OpenCode, add a custom provider under providers in ~/.config/opencode/opencode.json with the
@opencode/ai/providers/openai-compatible package, settings. set to https://api.batchwell.ai/v1, your Batchwell API key
read from BATCHWELL_API_KEY, and a Batchwell model id in its models map. Then set model to the provider key, a
slash and the model id. This is OpenCode's documented setup for an OpenAI-compatible provider, in OpenCode 2.
Create a key as described in the Quickstart and Authentication, then run:
It writes ~/.config/opencode/opencode.json and ~/.config/batchwell/api-key: Batchwell's provider and the model line in OpenCode's global config, and
your key in a file only you can read, which the config reads with {file:.. Every other key in opencode.
keeps its value. It reads your key from BATCHWELL_API_KEY, or asks for it on the terminal with the input hidden, and
checks it with GET /v1/models before it writes anything. Then start the agent with opencode.
Every file it changes is first copied to <file>.; to undo it, copy that backup back over the
file, or delete a file the command created. The command needs OpenCode 2 and jq, and runs on macOS, Linux and WSL;
see Connect an agent for every option.
Paste this prompt into the agent, or into any coding agent, and it sets itself up the same way. It names the variable that holds the key, never the key itself, and tells the agent to back up each file before it changes it.
The snippet is made for batchwell/kimi-k3.1-pro-uncensored, a model Batchwell serves today. Merge it into
~/.config/opencode/opencode.json, keeping your other keys. The key comes from the BATCHWELL_API_KEY environment
variable, so no key is written into the file. Create a key as described in the Quickstart and
Authentication.
OpenCode reads opencode. as JSON or JSONC. A project's own opencode. is merged over the global file, so a
project that sets model itself keeps its own choice there.
modelpicks the model as the provider key, a slash, and the model id.providersholds the provider, under the key the model reference uses before the slash.packageselects OpenCode's runtime for an OpenAI-compatible API, which calls the OpenAI Chat Completions API that Batchwell serves.settings.is the address OpenCode sends requests to. Set it to the Batchwell base URL, which ends inbaseURL /v1.settings.is the bearer key OpenCode sends.apiKey {env:BATCHWELL_API_KEY}reads it from the environment. The key starts withbw_live_orbw_test_.modelslists the model: the key is the model id fromGET /v1/models(see models) andnameis the label OpenCode shows.
OpenCode also accepts limit on each model, with the context and output token limits. Set them to the model's limits
from models.
That model lists these supported parameters: temperature, top_p, max_tokens, max_completion_tokens, stop, stream_options, tools, tool_choice, reasoning_effort. These are the model's tunable
parameters, not the whole request: Batchwell accepts only the fields listed under
accepted fields, including stream and stream_options. Any other field fails
with 400 unsupported_parameter, and the error message names the field.
Batchwell rejects n greater than 1, logprobs and top_logprobs with 400 unsupported_parameter, and accepts text content only (parts of
type image_url, input_audio or file get 400 unsupported_content). The full list is on
chat completions.
- Connect an agent compares the five agents and every option of the connect command.
- Routing covers fallback models and routing preferences.
- Errors lists every error code and what to do about it.
- Codex setup and Claude Code setup show the same model in two of the other coding agents.