API harnesses

API harnesses call an LLM API directly instead of running a coding agent. A step on an API harness sends one request (system prompt + your rendered prompt) and uses the response text as the step’s output.

No tools, no files

API steps can’t read the repository, edit files or run commands. They only see what you put in the prompt. Use them for triage, specs, summaries, classification, and reviewing text you include, such as {{steps.plan.output}} or a diff captured by an earlier shell step. For work on the code, use a CLI harness.

Two API harnesses are built in: Claude API and OpenAI-compatible API. To use several OpenAI-compatible providers at once (say OpenAI and a local Ollama), add more with Add harness and set Type to LLM API.

Editor fields

Field	Meaning
Provider	Anthropic (Claude API) or OpenAI-compatible (OpenAI, OpenRouter, Ollama, LM Studio…)
Base URL	Anthropic: leave empty for api.anthropic.com. OpenAI-compatible: the URL that /chat/completions is appended to, e.g. http://localhost:11434/v1. Empty means https://api.openai.com/v1.
API key	Stored encrypted with your OS keychain. Takes precedence over the environment variable.
…or environment variable	Name of the env var holding the key. Empty means ANTHROPIC_API_KEY (Anthropic) or OPENAI_API_KEY (OpenAI-compatible).
Max output tokens	Sent as max_tokens (Anthropic) or max_completion_tokens (OpenAI-compatible).
Models	The one-click choices in the step editor, with a ★ default. Fetch latest pulls the provider’s list.

The key lookup order is: API key on the harness → the named env var in the step’s environment (which includes Settings → Execution → Environment variables, project env and step env) → the same env var in the app’s own environment.

Detection and Fetch latest only see the app environment

The Harnesses page’s availability check and the Models / Fetch latest buttons use the app’s environment, which doesn’t include Settings → Environment variables. If your key is only set there, runs work but the card shows Set an API key or ANTHROPIC_API_KEY and fetching models fails. Paste the key into the harness’s API key field to get both.

Claude API

Setting	Default
Id	anthropic-api
Provider	Anthropic
Key env var	ANTHROPIC_API_KEY
Max output tokens	32000
Models	claude-opus-5, claude-opus-5-5, claude-fable-5-1, claude-sonnet-5, claude-haiku-4-5
Default model	claude-opus-5

How a step runs:

• Uses the official Anthropic SDK and streams the Messages API response. Text appears in the log line by line as it arrives.
• Sends the step’s system prompt as system and the rendered prompt as a single user message.
• Enables adaptive thinking (thinking: { type: "adaptive" }).
• If no key is found at all, the SDK falls back to its own credential lookup.
• Records input, output, cache-read and cache-write tokens. The API doesn’t return a price, so the step’s cost stays $0.
• Fails with The model declined this request (refusal) on a refusal stop reason, and with Response hit max_tokens if the answer was cut off. Raise Max output tokens in that case.

The Bug fix (fast lane) template uses the Claude API for its triage step.

Fetch latest lists models from the Anthropic Models API.

OpenAI-compatible API

Setting	Default
Id	openai-compatible
Base URL	https://api.openai.com/v1
Key env var	OPENAI_API_KEY
Max output tokens	16000
Models	gpt-5, gpt-5-mini
Default model	gpt-5

How a step runs:

• POST {base URL}/chat/completions with model, messages (a system message if the step has a system prompt, then the user prompt) and max_completion_tokens.
• Sends Authorization: Bearer <key> only if a key was found, so keyless local servers work.
• Not streamed: the whole answer is logged when it arrives.
• Uses choices[0].message.content as the output and records prompt_tokens / completion_tokens if the server returns usage. No cost.
• The request timeout is the step’s timeout, or 10 minutes if the step has none.

Setting Max output tokens to 0 leaves max_completion_tokens out of the request, which helps with servers that reject that field.

Fetch latest calls GET {base URL}/models (with the bearer key if any) and reads the data array.

Provider settings

OpenAI

Field	Value
Base URL	empty, or https://api.openai.com/v1
Key	an OpenAI API key, or OPENAI_API_KEY
Models	e.g. gpt-5, gpt-5-mini — or click Fetch latest

OpenRouter

Field	Value
Base URL	https://openrouter.ai/api/v1
Key	an OpenRouter key (set …or environment variable to e.g. OPENROUTER_API_KEY if you keep it in the environment)
Models	provider-prefixed ids such as anthropic/claude-sonnet-5; Fetch latest returns OpenRouter’s full catalog

Ollama

Field	Value
Base URL	http://localhost:11434/v1
Key	none needed
Models	the names you pulled, e.g. qwen3-coder:30b; Fetch latest lists local models

A localhost base URL counts as available without a key. See Local models with Ollama.

LM Studio

Field	Value
Base URL	http://localhost:1234/v1 (LM Studio’s local server default)
Key	none needed
Models	the loaded model’s id; Fetch latest lists them

vLLM

Field	Value
Base URL	http://<host>:8000/v1 (vLLM’s OpenAI-compatible server default port)
Key	only if you started vLLM with --api-key
Models	the served model name; Fetch latest lists it

A vLLM server on another machine isn’t localhost, so detection reports it unavailable until a key or key variable is set. Runs still work without one if the server doesn’t require it.

Adding a second API harness

1. Harnesses → Add harness.
2. Set Name (e.g. Ollama) and switch Type to LLM API.
3. Choose Provider, fill in Base URL, and either API key or …or environment variable.
4. Add model ids under Models (press Enter after each; click ★ to set the default), or click Fetch latest.
5. Save. The harness appears under LLM APIs and in every step’s Harness menu, marked (API).

Timeouts and cancellation

The step’s Timeout (minutes) applies to API calls too; a timed-out call fails with Timed out after N min. Cancelling a run aborts the in-flight request.