The biggest issue I have with this set-up is that my MCP tools don't seem to work with the --hybrid-tools

It can't find the tool which I guess is intended behaviour

Expected behavior, and worth documenting — --hybrid-tools rewrites the tool array with CC's exact Bash / Read / Write / Edit / Glob / Grep / WebSearch / WebFetch set to preserve the subscription fingerprint. Any MCP tool your OpenClaw config declared gets dropped in that swap, which is why "tool not found" fires when the model tries to call it.

Three options depending on what you care about most:

1. --preserve-tools (not --hybrid-tools) — dario forwards OpenClaw's full tool array to Anthropic verbatim, MCP tools included. Trade-off: you lose the CC tool fingerprint (the subscription sees your custom schemas, which is itself a fingerprint tell). MCP just works.

2. Default mode (no tool flag) — dario's built-in TOOL_MAP translates OpenClaw's standard tools (exec, process, web_search, etc.) to CC tool names, keeping the CC fingerprint for those. MCP tools aren't in the map, so they flow through as unmapped — the model can still call them by name, but they occupy CC tool slots and the fingerprint is mixed.

3. Split — run OpenClaw sessions that need MCP through a direct Anthropic key and reserve dario for the sessions where subscription routing matters. Not ideal but the cleanest when a single session needs both CC-fingerprint and rich MCP.

Short answer if you do not want to tune it: drop --hybrid-tools, add --preserve-tools. MCP will work, you pay the fingerprint cost.

On your other config points — contextWindow 200000 is correct for Claude 4 models (that is the model-level max, not a dario limit). maxTokens 32000/64000 in your OpenClaw JSON is the client-side cap; dario pins the wire value to 32000 now anyway (that is what real CC v2.1.114 sends), so the value OpenClaw declares has no effect on what Anthropic sees. You can leave it at 64000, will not change anything.

Your setup looks solid! A few observations and suggestions:

Context Window & Max Tokens:
Your contextWindow: 200000 is correct for both Opus and Sonnet (they support 200k context), but I'd actually verify the maxTokens values. Opus typically outputs up to 4096 tokens by default in the API, not 32000 — you might want to check the actual limits depending on your Dario version. Setting it too high could cause unexpected behavior.

Config Structure Question:
One thing I notice — you're setting claude-haiku-4-5 in the agents section but not defining it in your providers/dario/models array. Make sure all models listed in agents are actually declared above, otherwise OpenClaw might throw errors during fallback.

Hybrid-Tools Mode:
Running dario proxy --hybrid-tools is the right call for tool use. Just worth confirming: are you seeing the tools actually execute properly, or are they being passed through without execution?

What I'd Test:

• Try hitting the dario endpoint directly with curl to verify response times (curl -X POST http://localhost:3456/v1/messages ...)
• If you're doing heavy reasoning tasks, the 64k max tokens on Sonnet might hit limits faster than expected — consider lowering fallback to Haiku for non-critical calls to preserve quota

Quick Tip:
If you're managing multiple model providers and want to avoid having to manually maintain all these config entries, some folks spin up a lightweight proxy layer that auto-discovers available models. Not necessary for your current setup, but worth keeping in mind if it grows.

What kind of workloads are you running through this? That'd help tune the fallback strategy.

Quick correction on a couple of the numbers — dario pins outbound max_tokens to 32000 for both Opus and Sonnet as of v3.30.1 (that is what real CC v2.1.114+ sends, and dario matches the CC wire shape on that axis). The value OpenClaw declares in maxTokens in its config is client-side only and does not reach Anthropic, so maxTokens: 32000 or 64000 in the OpenClaw JSON is cosmetic — same outbound.

The 4096 number predates CC's thinking-token model. Opus 4.6 with effort: high (also pinned by dario v3.30.1+ to match CC 2.1.114+ wire shape) plus adaptive thinking routinely emits well above that; 32000 is the current cap on the wire.

Rest of the suggestions (direct curl test, Haiku fallback for non-critical calls) hold up.