Build an OpenClaw config that gives each job the right model: a primary and fallback, a cheap model for sub-agents and one for images. Start from a preset, see the monthly cost, then copy or download openclaw.json.
Free OpenClaw config generator. Builds the agents block of openclaw.json with primary, fallback, sub-agent and vision models, thinking level, concurrency and timeouts, plus a monthly cost estimate that includes thinking tokens.
Thinking level
A good default for most agents.
Parallel agents
Tasks running at once
Task timeout
Longest a main task may run
Sub-agent timeout
Longest a sub-task may run
Context sent per call
More context costs more on every step
Image upload limit
Megabytes per image
Your setup
~$10a month at 330 tasks
Primary about $10, sub-agents about $0.16. Includes thinking tokens at each model's everyday setting.
// OpenClaw agents config (JSON5). Merge into openclaw.json
// Profile: Solopreneur
// Generated: 2026-09-26 via dervity.com
{
// Set OPENROUTER_API_KEY in your environment
env: { OPENROUTER_API_KEY: "sk-or-..." },
agents: {
defaults: {
// Primary and fallback models
model: {
primary: "openrouter/openai/gpt-6-sol",
fallbacks: ["openrouter/deepseek/deepseek-v4.1-flash"],
},
// Model catalog (aliases for quick switching)
models: {
"openrouter/openai/gpt-6-sol": { alias: "GPT-6 Sol" },
"openrouter/deepseek/deepseek-v4.1-flash": { alias: "DeepSeek V4.1" },
"openrouter/openai/gpt-6-luna": { alias: "GPT-6 Luna" },
},
// Vision / image analysis model
imageModel: {
primary: "openrouter/deepseek/deepseek-v4.1-flash",
},
// Subagent settings (delegated tasks)
subagents: {
model: "openrouter/openai/gpt-6-luna",
maxConcurrent: 1,
runTimeoutSeconds: 600,
archiveAfterMinutes: 60,
},
// Agent defaults
thinkingDefault: "medium",
maxConcurrent: 2,
timeoutSeconds: 300,
mediaMaxMb: 5,
contextTokens: 200000,
},
},
}Written for OpenClaw. Adapt the keys for other agent frameworks.
Was this tool helpful?
An OpenClaw config: the agents block of openclaw.json, written in JSON5 so it can carry comments. It sets your primary and fallback models, a cheaper model for sub-agents, a vision model, the default thinking level, and concurrency and timeout limits. Pick a preset, change what you like, then copy or download it.
Merge the agents block into your openclaw.json, or download the file and use it as a starting point. OpenClaw checks the file against its schema when it loads, so a typo in a key shows up as an error. To confirm the model took, run openclaw config get agents.defaults.model.primary.
OpenRouter mode sends every model through OpenRouter, so one API key covers Anthropic, OpenAI, Google and most other labs, and model names get an openrouter/ prefix. Direct mode talks to each provider's own API. It skips the middle layer, but you need a key per provider.
How much the model reasons before it answers. Higher settings help with hard coding and planning tasks but take longer and bill more thinking tokens. Off skips extended thinking, which suits quick replies and simple lookups. Medium is a sensible default for most agents.
Sub-agents handle small delegated jobs like file lookups, searches and simple edits, and an agent can start many of them in one task. A fast, cheap model does that work well, so routing it away from your primary model cuts the bill without hurting the main answer.
From list prices and a typical agent workload, with each model's thinking tokens included at its everyday reasoning setting, the same method as our best-model-for pages. Long contexts, caching and how often sub-agents run will move your real bill, so treat it as a ballpark.
Budget for the lowest bill from models that still score well, Solopreneur for strong answers on a small bill, Developer for the top coding model with fast, cheap helpers, and Agency or team for top quality everywhere. Every preset is a starting point, and each model and limit can be changed.
Everything from model picks to CSV converters, with no account needed.