菜单
Configure OpenMolClaw's model provider: a local chat-completions endpoint, OpenRouter ZDR, or another explicit endpoint behind one ModelProvider interface. Use when wiring, swapping, or debugging the model the router and executor call.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
How to embed Ketcher in a local Flask + JavaScript page and bridge it to RDKit: read/write Molblock, convert pasted SMILES to Molblock server-side, and render validated structures into the local workspace. Use when building or extending the local structure-editor harness.
Conventions for drawing accurate 2D skeletal chemical structures in SVG. Covers zigzag backbones, wedge/dashed-wedge bonds, heteroatom labels, stereochemistry notation, and common pitfalls.
Programmatically build, modify, and manipulate chemical structures using RDKit without requiring LLM API calls. Use when you need to: (1) Retrieve common chemical scaffolds (benzene, pyridine, amino acids, etc.), (2) Build molecules from building blocks (chains, rings, functional groups), (3) Manipulate fragments (replace substructures, find disconnections, extract MCS), (4) Validate or canonicalize SMILES strings, (5) Generate molecular series or variants (homologous series, isomers, functional group scans), or (6) Perform any routine structure manipulation for local chat or workspace workflows.
The router-first agent pattern: user request → router → forced tool choice → executor → tool result, with JSON-schema validation, an execution trace, and deterministic fallback errors. Use when building or reasoning about the local agent harness loop so model text is never executed directly.
The local workspace JSON format and semantic-alias system ([m1], [r1], [label1]): how molecules, reactions, labels, and rendered assets are stored, referenced, and round-tripped to a single local JSON file with no database. Use when building or reasoning about local workspace state and object references.
| name | local-model-provider |
| description | Configure OpenMolClaw's model provider: a local chat-completions endpoint, OpenRouter ZDR, or another explicit endpoint behind one ModelProvider interface. Use when wiring, swapping, or debugging the model the router and executor call. |
OpenMolClaw calls a configured model provider through a single interface, so the router and executor never care which model answers. The default path is local, the recommended hosted path is OpenRouter ZDR, and any other chat-completions-compatible endpoint can be selected explicitly.
class ModelProvider:
def complete_with_tools(self, messages, tools, tool_choice):
...
Point at any local runtime that serves /v1/chat/completions:
model:
provider: local
model: olmo
endpoint: http://localhost:11434/v1
from openmolclaw.config import build_provider, load_config
provider = build_provider(load_config("examples/config.local.olmo.yaml"))
No API key is required by default.
model:
provider: openrouter
model: google/gemma-4-26b-a4b-it
base_url: https://openrouter.ai/api/v1
zdr: true
Set the key in the environment, never in the file:
export OPENROUTER_API_KEY=...
from openmolclaw.builtin_tools import build_default_registry
from openmolclaw.harness.router import Router
from openmolclaw.harness.executor import ToolExecutor
registry = build_default_registry()
router = Router(provider)
decision = router.route("Draw benzene.", registry.specs())
result = ToolExecutor(registry).execute_decision(decision)
Name any provider when you also provide the exact endpoint:
model:
provider: openai
model: gpt-4.1-mini
endpoint: https://api.openai.com/v1
api_key_env: OPENAI_API_KEY
For non-chat-completions APIs, implement complete_with_tools on any class and
pass the instance to Router. A host can wire its own backend behind this same
interface without touching harness code.
Maintained by the ChemIllusion team as part of OpenMolClaw.