OpenClaw Model Switch
Switch or repair the model used by the running OpenClaw instance. Prefer its validated native configuration commands; retain the bundled file tools for offline and older-version work.
Diagnose before you edit. Model failures on OpenClaw are usually NOT the model id — they are key routing (env hijack), provider-plugin restrictions, or endpoint/model mismatch. Changing the model id without checking these first is how a 5-minute fix becomes a 2-hour debugging session. The full trap catalog with discovery commands lives in references/troubleshooting-model-config.md — read it the moment anything errors.
Step 1 — Find the real config file(s)
Do NOT assume a hardcoded path. Candidate locations (check all, edit all that exist):
~/.openclaw/openclaw.json— the gateway's live config on most installs~/.kimi/kimi-claw/openclaw.json— Kimi Claw mirror, kept in sync on some installs~/.kimi_openclaw/openclaw.json— legacy desktop path
Confirm which one the gateway actually reads with openclaw gateway status (or its JSON
config paths). Select that installation's profile/config context before issuing commands.
If several files exist, verify which are maintained mirrors and keep those in sync;
existence alone does not prove they belong to the same instance. Reconcile an Ansible or
other deployment source too when it owns the default, so the next deployment preserves it.
Step 2 — Probe the endpoint + model BEFORE touching config
Never trust a relay's model listing (GET /v1/models on new-api style relays is frequently
incomplete — a model can be absent from the list yet serve fine). The only authority is a
real completion probe from the host that will run the bot:
Match the Gateway's effective credential and proxy route. An SSH login shell can inherit a dead proxy that the Gateway does not use; record that difference before interpreting a failed probe. Do not print keys or dump the process's full environment. Use the exact model variant requested by the user and verify its input modes, context/output limits and supported thinking settings from the provider's current documentation; do not infer them from the family name.
curl -sS -o /tmp/probe.json -w "HTTP %{http_code}\n" \
-X POST "<baseUrl>/v1/messages" \
-H "Authorization: Bearer <apiKey>" \
-H "Content-Type: application/json" \
-H "anthropic-version: 2023-06-01" \
-d '{"model":"<model-id>","max_tokens":16,"messages":[{"role":"user","content":"hi"}]}'
Expected: HTTP 200 and a content array in the body. 401 Invalid token with a token you
just verified works elsewhere → the wire key is being hijacked (see trap #1 in the
troubleshooting reference). 503 No available channel → the model is not served for this
token/group from this network (trap #3) — pick a served model or fix the relay, don't
blind-switch.
Step 3 — Switch the model
Read openclaw config set --help on the selected installation. On versions supporting batch
writes, prepare a private JSON batch containing the model definition and default assignment:
[
{"path":"models.providers.example.models","value":[{"id":"target-model","name":"Target model"}]},
{"path":"agents.defaults.model.primary","value":"example/target-model"}
]
Replace example with the existing provider id and target-model with the verified served
model. Fill the definition's reasoning, input and token limits from the verified model spec;
include agents.defaults.thinkingDefault only when the target requires a compatible change.
Preserve credentials, headers, plugins and unrelated model entries. Back up the active config
privately before applying the batch; do not commit either credentials or the live config.
openclaw config set --batch-file ./model-batch.json --merge --dry-run
openclaw config set --batch-file ./model-batch.json --merge
Use --merge to retain existing model definitions; use replacement only for an explicitly
authorized removal. Follow the CLI's application hint: hot reload, no change, or required
restart. The hint describes expected application, so complete Step 4 even after a hot-reload
hint. Restart only when required, after checking service ownership, in-flight work and recovery.
See native configuration commands.
Offline tools and older versions
Retain these helpers for exported/stopped configurations or a verified older CLI without the
native write interface. Back up the selected files and preserve maintained mirrors. A live
older-version repair must respect the same service and in-flight-work checks; --restart
is an explicit option, not the default completion step. This helper has no dry-run mode;
inspect a private exported copy first when the proposed file edits need a preview.
The legacy helper only supplies definitions for its known-model table and rewrites same-provider
entries in agents.defaults.models. For other targets, complete the verified definition and
restore any existing allowlist parameters in the offline candidate before adopting it. Use
explicit --config when discovery would include unverified mirrors; inspect the complete diff.
python3 scripts/switch-model.py <model-id>
# target a specific provider instead of the guessed one:
python3 scripts/switch-model.py k3 --provider kimi-relay
# explicit config path (skips discovery):
python3 scripts/switch-model.py k3 --config ~/.openclaw/openclaw.json
The script: discovers and backs up every candidate config to <config-dir>/config-backups/,
adds the model definition if known, sets agents.defaults.model.primary, syncs mirror
files, and restarts the gateway with --restart.
Step 4 — Verify end-to-end (mandatory)
A local/isolated turn is a provider precheck. For the live switch, run a new session through
the Gateway without --local and without a --model override, then read the result metadata:
openclaw agent --json --agent main --session-id verify-$(date +%s) -m "Reply with a short verification marker."
Require a successful command, a visible reply and the requested model in the actual execution.
For 2026.11.1, inspect result.meta.agentMeta.terminalReceipt: requested.model,
effective.model and effective.responseModel must match the intended variant, and rerouted
must be false. Also require result.meta.executionTrace.fallbackUsed=false. A successful
reply rescued by fallback does not validate the requested model. On older schemas, use the
equivalent execution metadata and Gateway logs; missing identity evidence remains unverified.
Report the tested boundary precisely. A Gateway turn does not test channel ingress or delivery. If delivery is part of the request, use an authorized target and inspect its send receipt; claim an inbound channel round trip only after observing a new message on that channel.
Common failures → read the troubleshooting reference
| Symptom | Most likely trap |
|---|---|
LLM error new_api_error: Invalid token, but the token works in curl |
Trap #1 — env KIMI_API_KEY hijacks the provider's wire key |
Thinking level "max" is not supported ... Use one of: off, on |
Trap #2 — kimi-provider plugin hardcodes binary thinking; bypass with a custom provider |
Thinking level ... Use one of: off, minimal, low, medium, high |
Trap #2 variant — anthropic-messages base profile; unlock via params.canonicalModelId |
503 No available channel for model X under group default |
Trap #3 — model not served for this group/network; listing ≠ availability |
| Edit saved + gateway restarted, nothing changed | Trap #5 — edited the wrong file / mirror not synced |
Safety rules
- Always backup before editing (the script does this; manual edits: copy to
config-backups/first) - Preserve existing
apiKey,headers, plugin configs, andenvblocks — retype only the fields you mean to change - Validate JSON after manual edits:
python3 -m json.tool openclaw.json > /dev/null - Do not commit config files containing API keys to version control
- If verification fails, preserve its result and reconcile or restore the owned config through the matching native interface. Restore an offline file backup only to its verified stopped owner; do not overwrite a running managed store or newer concurrent changes.
Resources
- scripts/switch-model.py — model switcher with config discovery, backup, mirror sync, and restart
- references/kimi-models.md — known model specs (k3, k2p6, kimi-k2.7-code) and config snippets
- references/troubleshooting-model-config.md — the trap catalog: env key hijack, plugin binary thinking, canonicalModelId, relay availability, config discovery. Read on any error.
