Model selection

[lusty cloud]

What models should I run for my main agent?
I have access via codex subscription and Anthropic sub so pricing/usage is a concern, I heard sonnet 3.5 is good?

[frank yoke]

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

[worldly olive]

If pricing matters, I’d optimize for tool reliability first on your main agent, then push “cheap” onto background stuff.

• Main agent: a “Sonnet-class” Anthropic model is usually the sweet spot (good reasoning + generally solid tool calling). So yes — the idea that “Sonnet 3.5 is good” has been true for a lot of people, but model lineups change fast, so I’d sanity-check what folks are currently running this week in this channel.
• Coding-heavy work: if you’ve got a Codex subscription, it can be great to use that for coding tasks specifically while keeping the main agent on the model that behaves best with tools + browsing + long, messy context.
• Cost control: run a cheaper model for background/automation (heartbeats, cron reminders, lightweight checks), and reserve the pricier model for interactive “I need it to actually succeed” runs.

Two quick questions so the recommendation matches your use:

1. Is your main agent mostly coding, or mostly browsing / automation / chat + planning?
2. Do you care more about speed or accuracy/tool-success rate?

Docs worth skimming (esp. the community-notes section):
https://docs.openclaw.ai/concepts/models
https://docs.openclaw.ai/concepts/models#model-discussions-community-notes

[lusty cloud]

I wouldn’t even know how to answer you, write 5 questions for my main agent and I’ll reply with his respond

[worldly olive]

Paste these 5 questions to your main agent and reply here with its answers (and any extra notes it adds):

1. What do you actually do most of the time?
   Pick up to 3: (a) coding/debugging (b) writing/brainstorming (c) planning/project mgmt (d) web browsing/research (e) automation/tool use (f) chat/companion

2. Which tools do you use regularly inside OpenClaw?
   List any: browser, exec, filesystem (read/write/edit), messaging, cron, nodes/camera/screen, etc.
   Also: do you need “tool calling never breaks” more than “best reasoning”?

3. What are your hard constraints?

   • Max budget per day/week (roughly)
   • Preferred latency: fast (<10s) / medium / don’t care
   • Context size needs: small / medium / huge (lots of logs + docs)

4. What failure mode hurts you most?
   Choose one: (a) wrong answer (b) tool-call mistakes (c) slow responses (d) too expensive (e) gets stuck / doesn’t finish

5. Given I have access to both Anthropic and Codex, propose a 2–3 tier model plan (Main / Cheap background / “Big gun”), and say why for each tier.
   Include: what to run for daily use vs. what to switch to when things get messy.

Reply with the 5 answers and I’ll map that to a concrete “main + backup + cheap tasks” recommendation.

[lusty cloud]

How do I switch out my google api key

[worldly olive]

For OpenClaw, the Google Gemini “API key” is the env var GEMINI_API_KEY (provider google).

To switch/rotate it:

1. Update the key where the Gateway runs

• If you run the Gateway as a daemon/service: edit ~/.openclaw/.env on the gateway host and replace the value, e.g.
  • GEMINI_API_KEY=...new key...
• If you’re on Fly/Render/Docker/systemd/etc: update the service secret/env var there instead.

Note: OpenClaw loads .env (CWD + ~/.openclaw/.env) only for missing vars—so if GEMINI_API_KEY is already set in the service environment, changing the file won’t matter until you change/remove the existing env var.

2. Restart the Gateway (so it picks up the new env var)

3. Verify

• Run: openclaw models status (optionally --json / --probe-provider google)

Docs: https://docs.openclaw.ai/concepts/model-providers (Google Gemini section) + https://docs.openclaw.ai/help/environment