@@ -0,0 +1,42 @@
+# Use Claude Code with GLM, DeepSeek or Kimi
+
+ThinkWatch Lite connects Claude Code to a gateway on the same computer, which forwards each request to an upstream that serves the model asked for. A GLM, DeepSeek or Kimi model is either named in Claude Code, with `/model` or the model variables in `settings.json`, or reached by a routing rule that rewrites Claude's model names to it. All three providers offer an Anthropic-compatible address, so Claude Code's requests reach them without format conversion.
+
+## Before you start
+
+- ThinkWatch Lite, [installed](/docs/lite/install/), and Claude Code, run at least once.
+- An API key from the provider. For GLM, a Z.ai or BigModel account with a GLM Coding Plan can sign in from the app instead.
+- A Claude Pro or Max sign-in cannot serve as an upstream; once connected, Claude Code uses its gateway key instead.
+
+## Steps
+
+1. On the Upstreams page, choose **New upstream** and pick a **Service**:
+   - **DeepSeek** fills in `https://api.deepseek.com/anthropic` and the protocol. Enter the **API key**.
+   - GLM with a key: **Custom**, with `https://open.bigmodel.cn/api/anthropic` or `https://api.z.ai/api/anthropic` as the **Base URL**, **Protocol** set to **Anthropic Messages**, and the **API key**.
+   - GLM with an account: **Z.ai / BigModel account**. Select the **Account service**, tick **Acknowledge the notes above and continue signing in**, choose **Sign in** and authorize in the browser. The app creates an API key named `thinkwatch` on the account and saves the upstream.
+   - Kimi: **Custom**, with the Anthropic-compatible base URL from Kimi's documentation and **Protocol** set to **Anthropic Messages**. For Kimi For Coding, also turn on **Forward client identity**.
+
+   **Check connection** verifies the address and key and fetches the model list at no cost. Choose **Next** twice, then **Create**.
+2. On the Clients page, choose **Connect…** on the Claude Code row. The dialog lists the fields that change in `~/.claude/settings.json`: `env.ANTHROPIC_BASE_URL`, `env.ANTHROPIC_AUTH_TOKEN` (a new key named `claude-code`) and `env.CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY`. Choose **Connect**.
+3. Choose how the model is selected:
+   - **By name.** In Claude Code, `/model <model ID>` switches to the model, and `claude --model <model ID>` starts with it. To map Claude Code's model aliases to it, add these to the `env` block of `~/.claude/settings.json`; the Haiku one also runs background tasks.
+
+     ```json
+     "ANTHROPIC_DEFAULT_OPUS_MODEL": "<model ID>",
+     "ANTHROPIC_DEFAULT_SONNET_MODEL": "<model ID>",
+     "ANTHROPIC_DEFAULT_HAIKU_MODEL": "<model ID>"
+     ```
+
+     No routing rule is needed; the default route skips upstreams whose model list does not contain it.
+   - **By rewriting.** On the Routing page, open the `default` route and choose **Add rule**. With **Add condition**, add **Model** `claude-*`, and **Key** `claude-code` to leave other clients alone. Set **On match** to **Forward**, **Forward to** to the upstream, and **Change model to** under **Parameter rewrites** to the model ID. Choose **Add**, then **Save**. Claude Code keeps showing Claude's names.
+
+## Notes
+
+- **Format conversion.** Claude Code sends Anthropic Messages; with that protocol the request goes out unchanged, apart from identity fields such as `metadata.user_id`, which the gateway removes. Only an upstream in another format, such as an OpenAI-compatible address set to OpenAI Chat Completions, gets a converted request: Traffic marks it **Converted** and its details list any dropped fields. Web search, a server-side tool, cannot be converted and is not sent to such an upstream. **Auto-detect** does not recognize these addresses and forwards requests in the client's own format, which suits Claude Code but not Codex.
+- **Forward client identity** is off by default, so upstreams see ThinkWatch's User-Agent and no client identity. Kimi For Coding, Bailian Coding Plan and similar upstreams accept only certain clients; with the switch on, they receive Claude Code's own User-Agent, identity headers such as `x-app` and the identity fields in the body, unaltered.
+- **GLM Coding Plan.** An upstream on `api.z.ai` or `open.bigmodel.cn`, signed in or added with a key, shows its 5-hour and weekly limits, and the credits left on a plan billed in credits, in the Quota / billing column and in the menu bar or tray menu.
+- **The /model list** shows gateway models only when their names contain `claude` or `anthropic`. Other models are typed by name, or added as one entry with `ANTHROPIC_CUSTOM_MODEL_OPTION`.
+- **Cost.** A rewritten request is priced as the model actually sent.
+- **Restore…** puts back only the fields the app wrote; model variables added by hand stay in `settings.json`.
+
+Related: [Features](/docs/lite/features/), [Use Codex with Claude, Gemini or a Chat Completions-only relay](/docs/lite/codex-other-models/), [Use Claude Desktop with third-party models](/docs/lite/claude-desktop-third-party-models/).
@@ -0,0 +1,36 @@
+# Use Claude Desktop with third-party models
+
+ThinkWatch Lite connects Claude Desktop through its official third-party inference mode, with the gateway on the same computer as the inference provider. Claude Desktop accepts only model names that look like Claude's, so a routing rule rewrites those names to the model that serves the request, from GLM, DeepSeek, Kimi or any other upstream. A Claude Desktop managed by an organization is left unchanged.
+
+## Before you start
+
+- ThinkWatch Lite, [installed](/docs/lite/install/), and Claude Desktop on macOS or Windows, opened at least once. The third-party inference mode needs no Anthropic account.
+- An upstream for the model, added on the Upstreams page. [Use Claude Code with GLM, DeepSeek or Kimi](/docs/lite/claude-code-other-models/) shows how for those three.
+
+## Steps
+
+1. On the Clients page, choose **Connect…** on the Claude Desktop row. The dialog lists four files, under `~/Library/Application Support` on macOS; on Windows, `Claude-3p` is under `%LOCALAPPDATA%` and `Claude` under `%APPDATA%`.
+
+   | File | Change |
+   |---|---|
+   | `Claude-3p/configLibrary/7477a7c4-1ce0-4d3a-9b1e-7477a7c40001.json` | A configuration named ThinkWatch: `inferenceProvider` set to `gateway`, the gateway address, the key with the `x-api-key` scheme, `chatTabEnabled` and the model list `inferenceModels` |
+   | `Claude-3p/configLibrary/_meta.json` | Adds the ThinkWatch entry and points `appliedId` at it; other configurations stay |
+   | `Claude-3p/claude_desktop_config.json` | `deploymentMode` set to `3p`, nothing else |
+   | `Claude/claude_desktop_config.json` | `deploymentMode` set to `3p`; the MCP servers in it stay |
+
+   `inferenceModels` receives the gateway's models whose names look like Claude's: `claude-` followed by `sonnet`, `opus`, `haiku` or `fable` and a version. When there are none, the notes say that `claude-sonnet-5` is written and give an example rule. Choose **Connect**.
+2. On the Routing page, open the `default` route and choose **Add rule**. With **Add condition**, add **Model** `claude-*` and **Key** `claude-desktop`, the key named in the connect dialog. Set **On match** to **Forward** and **Forward to** to the upstream, and under **Parameter rewrites** enter the model ID in **Change model to**. Choose **Add**, then **Save**.
+3. Quit Claude Desktop completely and open it again. If the sign-in page appears, choose to continue with the gateway there; this happens once.
+
+## Notes
+
+- **Managed by an organization.** A managed configuration overrides everything set on the computer: on macOS a `com.anthropic.claudefordesktop.plist` under `/Library/Managed Preferences`, on Windows the registry key `SOFTWARE\Policies\Claude` under `HKLM` or `HKCU`. The Clients page then offers no **Connect…**, and the details say "Claude Desktop on this computer is managed by an organization".
+- **Model list.** `inferenceModels` is written when connecting and is not offered for update afterwards. After the gateway's models change, **Restore…** and connect again to refresh it.
+- **Cloud providers.** When Claude Desktop uses Amazon Bedrock, Google Cloud Agent Platform or Microsoft Foundry through another configuration, the dialog says so: the ThinkWatch configuration is used while connected, and restoring switches back. For Bedrock it also offers **New Bedrock upstream…** with the previous settings.
+- **History and web search.** Conversations in this mode are kept apart from the existing ones. Web search does not work through the gateway and needs its own setup.
+- **Restore…** sets `deploymentMode` back in both files, removes the ThinkWatch entry from `_meta.json`, points `appliedId` back at the configuration used before if it still exists, and deletes the ThinkWatch configuration.
+- **Checks.** The details report "Another configuration is in use in Claude Desktop" when another configuration has been applied in the app, and "Claude Desktop may still open in its usual mode" when `deploymentMode` is not `3p`.
+- **By hand.** The same setup can be made in Claude Desktop: turn on Help → Troubleshooting → Enable Developer Mode, then open Developer → Configure Third-Party Inference. Choose the gateway provider, enter the gateway address and the key, set the authentication scheme to `x-api-key`, and click Apply Changes.
+- **Cost.** A rewritten request is priced as the model actually sent.
+
+Related: [Features](/docs/lite/features/), [Use Claude Code with GLM, DeepSeek or Kimi](/docs/lite/claude-code-other-models/), [Install and update](/docs/lite/install/).
@@ -0,0 +1,38 @@
+# Use Codex with Claude, Gemini or a Chat Completions-only relay
+
+ThinkWatch Lite connects Codex to a gateway on the same computer that accepts the OpenAI Responses API, the only format Codex uses with a custom provider. The gateway converts each request and its answer between Responses and the upstream's format: Anthropic Messages for Claude, Google Gemini, or OpenAI Chat Completions for a relay that offers nothing else. The model is named in Codex's configuration or set by a routing rule.
+
+## Before you start
+
+- ThinkWatch Lite, [installed](/docs/lite/install/).
+- Codex, either the Codex CLI or the Codex in the ChatGPT desktop app, run at least once.
+- An API key for Anthropic, Google Gemini or the relay.
+
+## Steps
+
+1. On the Upstreams page, choose **New upstream** and pick a **Service**: **Anthropic** for Claude, or **Google Gemini**; each fills in the address and protocol. For a relay, choose **Custom**, enter its **Base URL** without an endpoint path such as `/chat/completions`, and set **Protocol** to **OpenAI Chat Completions**. Enter the **API key**, choose **Check connection**, then **Next**. If the relay does not list its models, enter them one per line under **Manual list**. Choose **Next**, then **Create**.
+2. On the Clients page, choose **Connect…** on the Codex row. The dialog shows the change to `~/.codex/config.toml`:
+
+   | Field | Value |
+   |---|---|
+   | `model_provider` | `thinkwatch` |
+   | `model_providers.thinkwatch.base_url` | The gateway address with `/v1`, by default `http://127.0.0.1:8788/v1` |
+   | `model_providers.thinkwatch.wire_api` | `responses` |
+   | `model_providers.thinkwatch.experimental_bearer_token` | A new key named `codex` |
+   | `model_providers.thinkwatch.http_headers` | `X-ThinkWatch-Client = "codex"` |
+   | `name`, `requires_openai_auth`, `supports_websockets` in the same table | `ThinkWatch`, `false`, `false` |
+
+   Choose **Connect**, then reopen the terminal. The ChatGPT desktop app reads the same file and picks up the change after a restart.
+3. Name the model. Connecting leaves `model` as it was, so Codex keeps asking for the model it used before. Either:
+   - set `model = "<model ID>"` at the top of `~/.codex/config.toml`, before the first `[section]`, or pass `-c model=<model ID>` for a single run; or
+   - keep Codex's model and rewrite it: on the Routing page, open the `default` route and choose **Add rule**, add the conditions **Model** `gpt-*` and **Key** `codex`, set **On match** to **Forward** and **Forward to** to the upstream, and under **Parameter rewrites** enter the model ID in **Change model to**. Choose **Add**, then **Save**.
+
+## Notes
+
+- **Codex's model table.** Codex carries metadata for its own models, such as the context window, inside the program. A model it does not know, such as a Claude or Gemini model, runs on fallback metadata with a 272,000-token context window, and Codex warns: "Model metadata for `<model>` not found. Defaulting to fallback metadata; this can degrade performance and cause issues." For a model with a smaller window, `model_context_window` in `config.toml` sets the window Codex assumes. The gateway's model list does not appear in Codex's model picker, as Codex expects a catalog in its own format.
+- **Conversion.** Requests and streamed answers are converted in both directions. Traffic marks such requests **Converted**; fields the target format cannot carry are dropped and listed in the request details. Server-side tools such as web search run only at the provider they belong to and are dropped in conversion. Codex accepts only `responses` for `wire_api`, so this conversion is what makes a Chat Completions-only relay usable.
+- **Credentials.** With `requires_openai_auth = false`, Codex authenticates to the gateway with its own key and sends no OpenAI key or ChatGPT token. A ChatGPT account signed in from the app can still serve the OpenAI models next to Claude or Gemini: each request goes to an upstream that lists the model asked for.
+- **Sessions.** Codex lists sessions started before and after connecting separately. `codex resume <session ID> -c model_provider=thinkwatch` continues an earlier session through the gateway. After **Restore…**, sessions started while connected can still be opened, and go straight to OpenAI.
+- **Cost.** A rewritten request is priced as the model actually sent.
+
+Related: [Features](/docs/lite/features/), [Use Claude Code with GLM, DeepSeek or Kimi](/docs/lite/claude-code-other-models/), [Install and update](/docs/lite/install/).