feat(llm, ui): Add native keyless Local LLM model support (Ollama, vLLM, LM Studio, LocalAI) and Office UI Settings

README.md

@@ -611,29 +611,27 @@ The parent manager receives a capability summary compiled from the covered roles
 | `.opc/config/company_orgs/org_<id>_config.yaml` | Saved custom company architectures used by Company Mode. |
 | `.opc/config/org_index.yaml` | Active saved company architecture selector. |
 
-### LLM Keys
+### LLM Keys & Local Model Support
 
-After `opc init`, edit `.opc/config/llm_config.yaml` in the repo-local OPC home. If you set `OPC_HOME`, edit `$OPC_HOME/config/llm_config.yaml` instead.
-
-The template leaves secrets empty. Write your key directly into the file:
+After `opc init`, edit `.opc/config/llm_config.yaml` in the repo-local OPC home (or `$OPC_HOME/config/llm_config.yaml`). OpenOPC supports both Cloud LLMs (OpenAI, Anthropic, OpenRouter) and keyless **Local LLMs** (Ollama, vLLM, LM Studio, LocalAI, Llama.cpp, TGI):
 
 ```yaml
 llm:
-  default_model: "openai/gpt-5.4"
-  api_base: "https://openrouter.ai/api/v1"
-  api_key: "sk-or-v1-..."   # your OpenRouter (or other provider) API key
-
-  max_tokens: 32768         # max output tokens per request; lower it if your
-                            # model's output cap is smaller
-  # context_window: 128000  # total input window. Usually auto-detected via
-                            # litellm; unmapped models fall back to 128000.
-                            # Uncomment and set only when the fallback is
-                            # wrong for your model.
+  default_model: "ollama/llama3.3"         # or "vllm/meta-llama-3.1-8b-instruct", "openai/gpt-4o"
+  api_base: "http://localhost:11434"        # auto-resolved for local model prefixes
+  api_key: ""                               # keyless for local self-hosted nodes!
+
+  max_tokens: 32768         # max output tokens per request
+  is_local: true            # optional explicit local execution flag
 ```
 
+You can also configure LLM settings directly from the browser using the **LLM & Local Model Settings Modal (`🤖` button)** in the top header of Office UI.
+
+See [`docs/LOCAL_MODELS.md`](docs/LOCAL_MODELS.md) for detailed setup guides on Ollama, vLLM, LM Studio, and LocalAI.
+
 Then verify with `opc status`.
 
-If you prefer not to store the key in the file, leave `api_key` empty and set `api_key_env` to the name of an environment variable that holds it (e.g. `api_key_env: "OPENROUTER_API_KEY"`).
+If you prefer not to store cloud keys in the file, leave `api_key` empty and set `api_key_env` to the name of an environment variable that holds it (e.g. `api_key_env: "OPENROUTER_API_KEY"`).
 
 ### Approval & Agent Permissions
 

README.zh-CN.md

@@ -611,28 +611,27 @@ opc exec --mode task --agent jiuwenswarm "由一个 Team 完成这个任务"
 | `.opc/config/company_orgs/org_<id>_config.yaml` | Company 模式使用的自定义公司架构。 |
 | `.opc/config/org_index.yaml` | 当前生效的已保存公司架构选择器。 |
 
-### LLM 密钥
+### LLM 密钥与本地模型支持
 
-运行 `opc init` 后，编辑仓库本地 OPC home 中的 `.opc/config/llm_config.yaml`。如果设置了 `OPC_HOME`，则改为编辑 `$OPC_HOME/config/llm_config.yaml`。
-
-模板中的密钥留空。直接把 key 写入文件：
+运行 `opc init` 后，编辑仓库本地 OPC home 中的 `.opc/config/llm_config.yaml`（或 `$OPC_HOME/config/llm_config.yaml`）。OpenOPC 同时支持云端 LLM（OpenAI、Anthropic、OpenRouter）与无 API Key 的**本地模型**（Ollama、vLLM、LM Studio、LocalAI、Llama.cpp、TGI）：
 
 ```yaml
 llm:
-  default_model: "openai/gpt-5.4"
-  api_base: "https://openrouter.ai/api/v1"
-  api_key: "sk-or-v1-..."   # 你的 OpenRouter（或其他提供方）API key
-
-  max_tokens: 32768         # 每次请求的最大输出 token；如果你的模型
-                            # 输出上限更小，请调低
-  # context_window: 128000  # 总输入窗口。通常由 litellm 自动检测；
-                            # 未收录的模型回退为 128000。仅当回退值
-                            # 不适合你的模型时才取消注释并设置。
+  default_model: "ollama/llama3.3"         # 或 "vllm/meta-llama-3.1-8b-instruct"、"openai/gpt-4o"
+  api_base: "http://localhost:11434"        # 本地模型前缀将自动解析默认 Base URL
+  api_key: ""                               # 本地节点免 API Key 运行！
+
+  max_tokens: 32768         # 每次请求的最大输出 token
+  is_local: true            # 可选的显式本地运行标志
 ```
 
+你也可以直接在浏览器 Office UI 顶栏点击 **LLM & 本地模型设置弹窗（`🤖` 按钮）** 进行可视配置。
+
+详细配置指南请参见 [`docs/LOCAL_MODELS.md`](docs/LOCAL_MODELS.md)。
+
 然后用 `opc status` 验证。
 
-如果不想把密钥存在文件里，可以将 `api_key` 留空，并把 `api_key_env` 设置为持有密钥的环境变量名（例如 `api_key_env: "OPENROUTER_API_KEY"`）。
+如果不想把云端密钥存在文件里，可以将 `api_key` 留空，并把 `api_key_env` 设置为持有密钥的环境变量名（例如 `api_key_env: "OPENROUTER_API_KEY"`）。
 
 ### 审批与 Agent 权限
 

docs/LOCAL_MODELS.md

@@ -0,0 +1,64 @@
+# Running OpenOPC with Local LLM Models
+
+OpenOPC natively supports self-hosted local LLMs (Ollama, vLLM, LM Studio, LocalAI, Llama.cpp, TGI) without requiring remote API keys.
+
+---
+
+## 🚀 Supported Local Providers & Formats
+
+| Local Provider | Default Port | Model Prefix Format | Example Model Identifier |
+|---|---|---|---|
+| **Ollama** | `http://localhost:11434` | `ollama/<model>` | `ollama/llama3.3`, `ollama/qwen2.5-coder` |
+| **vLLM** | `http://localhost:8000/v1` | `vllm/<model>` or `openai/<model>` | `vllm/meta-llama-3.1-8b-instruct` |
+| **LM Studio** | `http://localhost:1234/v1` | `openai/<model>` | `openai/deepseek-r1-distill-qwen-14b` |
+| **LocalAI / Llama.cpp** | `http://localhost:8080/v1` | `openai/<model>` | `openai/starcoder2-15b` |
+| **Custom Local OpenAI Server** | `http://<host>:<port>/v1` | `openai/<model>` | `openai/custom-model` |
+
+> ⚠️ **Important Requirement**: Native OpenOPC agents require a local model that supports **tool calling / function calling** (such as `ollama/llama3.3` or `ollama/qwen2.5-coder`) to execute multi-step work items cleanly.
+
+---
+
+## ⚡ Quick Start: Running with Ollama
+
+1. **Start Ollama** locally and pull your model:
+   ```bash
+   ollama run llama3.3
+   ```
+
+2. **Configure OpenOPC `llm_config.yaml`**:
+   Edit `~/.opc/config/llm_config.yaml`:
+   ```yaml
+   llm:
+     default_model: "ollama/llama3.3"
+     api_base: "http://localhost:11434"
+     is_local: true
+   ```
+
+3. **Launch OpenOPC Session**:
+   ```bash
+   opc chat "Build a REST API in Python"
+   ```
+
+---
+
+## 🖥️ Running with LM Studio / LocalAI / OpenAI-Compatible Servers
+
+For OpenAI-compatible local servers (like LM Studio or LocalAI), use `openai/<model>` together with `api_base`:
+
+```yaml
+llm:
+  default_model: "openai/deepseek-r1-distill-qwen-14b"
+  api_base: "http://localhost:1234/v1"
+  is_local: true
+```
+
+---
+
+## 🌐 Office UI Settings Modal Integration
+
+Settings can be changed dynamically from the browser using the **LLM & Local Model Settings Modal (`🤖` button)** in the top header of Office UI:
+
+1. Click the **`🤖` button** in the header.
+2. Select your provider (Ollama, vLLM, LM Studio, LocalAI, OpenAI, Anthropic).
+3. Enter model identifier and API base URL.
+4. Click **Save & Apply Settings**. Changes are saved atomically to `.opc/config/llm_config.yaml` and reinitialize the runtime LLMProvider instantly.

opc/core/config.py

@@ -281,6 +281,8 @@ class LLMConfig(BaseModel):
     api_base: str = ""
     api_key: str = ""
     api_key_env: str = ""
+    provider: str = ""
+    is_local: bool = False
     routing: dict[str, str] = Field(default_factory=dict)
     fallback: dict[str, Any] = Field(default_factory=dict)
     temperature: float = 0.3
@@ -1739,6 +1741,55 @@ def load_active_company_org_payload(config_dir: Path) -> tuple[dict[str, Any], P
     return _validate_company_org_payload(path, payload), path
 
 
+class OPCConfig(BaseModel):
+    system: SystemConfig = Field(default_factory=SystemConfig)
+    llm: LLMConfig = Field(default_factory=LLMConfig)
+    agents: AgentsConfig = Field(default_factory=AgentsConfig)
+    org: OrgConfig = Field(default_factory=OrgConfig)
+    channels: ChannelsConfig = Field(default_factory=ChannelsConfig)
+    autonomy: AutonomyConfig = Field(default_factory=AutonomyConfig)
+    capabilities: CapabilityConfig = Field(default_factory=CapabilityConfig)
+    _trusted_workspace: Path | None = PrivateAttr(default=None)
+    _trusted_config_dir: Path | None = PrivateAttr(default=None)
+
+    def bind_workspace_trust(self, workspace: Path, config_dir: Path) -> None:
+        """Attach the project trust provenance used for pre-sink rechecks."""
+
+        self._trusted_workspace = Path(workspace).resolve(strict=False)
+        self._trusted_config_dir = Path(config_dir).resolve(strict=False)
+
+    def require_workspace_trust(self, *, include_effective: bool = False) -> None:
+        """Revalidate bound project authority immediately before runtime use.
+
+        Normal runtime checks bind the source files while allowing explicit
+        host-side CLI overrides such as ``--model`` and ``--debug``.  Loading
+        and trust-grant flows additionally validate the normalized declaration.
+        """
+
+        if self._trusted_workspace is None or self._trusted_config_dir is None:
+            return
+        from opc.core.workspace_trust import WorkspaceTrustStore
+
+        WorkspaceTrustStore().require(
+            self._trusted_workspace,
+            self._trusted_config_dir,
+            self if include_effective else None,
+        )
+
+def resolve_config_dir(path: Path | None = None) -> Path:
+    """Resolve target config directory (canonical: opc_home / 'config')."""
+    if path is None:
+        return get_opc_home() / "config"
+    p = Path(path)
+    if p.name == "config":
+        return p
+    if (p / "config").is_dir():
+        return p / "config"
+    if not (p / "llm_config.yaml").exists() and not (p / "system_config.yaml").exists():
+        return p / "config"
+    return p
+
+
 class OPCConfig(BaseModel):
     system: SystemConfig = Field(default_factory=SystemConfig)
     llm: LLMConfig = Field(default_factory=LLMConfig)
@@ -1781,9 +1832,7 @@ def load(
         *,
         trusted_source: bool = False,
     ) -> "OPCConfig":
-        if config_dir is None:
-            config_dir = get_opc_home() / "config"
-        config_dir = Path(config_dir)
+        config_dir = resolve_config_dir(config_dir)
         trusted_workspace: Path | None = None
         trust_store: Any | None = None
 
@@ -1872,9 +1921,16 @@ def load(
             config.bind_workspace_trust(trusted_workspace, config_dir)
         return config
 
+    def save_llm_config(self, config_dir: Path | None = None) -> Path:
+        """Atomically save ONLY llm_config.yaml under canonical config_dir."""
+        target_dir = resolve_config_dir(config_dir)
+        target_dir.mkdir(parents=True, exist_ok=True)
+        llm_path = target_dir / "llm_config.yaml"
+        _atomic_write_yaml(llm_path, {"llm": self.llm.model_dump()})
+        return llm_path
+
     def save(self, config_dir: Path | None = None) -> None:
-        if config_dir is None:
-            config_dir = get_opc_home() / "config"
+        config_dir = resolve_config_dir(config_dir)
         config_dir.mkdir(parents=True, exist_ok=True)
 
         system_path = config_dir / "system_config.yaml"

opc/engine.py

@@ -999,6 +999,30 @@ async def initialize(self) -> None:
             await self._recover_interaction_consumers()
             await self._recover_company_work_item_gate_continuations()
         logger.info("OPC Engine initialized successfully")
+
+    def reconfigure_llm(self, new_config: LLMConfig) -> LLMProvider:
+        """Consistently reconfigure LLMProvider for the engine and all dependent consumers."""
+        self.config.llm = new_config
+        new_llm = LLMProvider(new_config, opc_home=self.opc_home)
+        self.llm = new_llm
+
+        if getattr(self, "history_compactor", None) is not None:
+            self.history_compactor.llm = new_llm
+        if getattr(self, "communication", None) is not None:
+            self.communication.llm = new_llm
+        if getattr(self, "approval_engine", None) is not None:
+            self.approval_engine.llm = new_llm
+        if getattr(self, "secretary", None) is not None:
+            self.secretary.llm = new_llm
+        if getattr(self, "company_runtime_spec_builder", None) is not None:
+            self.company_runtime_spec_builder.llm = new_llm
+        if getattr(self, "company_recruiter", None) is not None:
+            self.company_recruiter.llm = new_llm
+        if getattr(self, "company_executor", None) is not None:
+            self.company_executor.llm = new_llm
+        if getattr(self, "task_router", None) is not None:
+            self.task_router.llm = new_llm
+        return new_llm
         if reconciled:
             logger.warning(
                 "Reconciled {} interrupted task(s) for project {} during startup",