converse-mcp-server 4.4.0 → 4.5.0

package/.env.example

@@ -47,6 +47,9 @@ DEEPSEEK_API_KEY=your_deepseek_api_key_here
 # append :online to a slug to opt into web search (adds a real per-request cost).
 OPENROUTER_API_KEY=your_openrouter_api_key_here
 
+# Get your abliteration.ai API key (ak_...) from: https://abliteration.ai/console/api-keys
+ABLITERATION_API_KEY=your_abliteration_api_key_here
+
 # Optional: referer and title for OpenRouter ranking credit (both optional; omitting is valid)
 # OPENROUTER_REFERER=https://github.com/FallDownTheSystem/converse
 # OPENROUTER_TITLE=Converse
@@ -102,6 +105,7 @@ TYPESAFE_API_KEY=your_typesafe_api_key_here
 # MISTRAL_DEFAULT_MODEL=mistral-medium-3-5
 # DEEPSEEK_DEFAULT_MODEL=deepseek-v4-pro
 # OPENROUTER_DEFAULT_MODEL=z-ai/glm-5.2  # any vendor/model slug
+# ABLITERATION_DEFAULT_MODEL=abliterated-model-large-v2
 
 # ============================================
 # Server Configuration

package/README.md

@@ -2,7 +2,7 @@
 
 [![npm version](https://img.shields.io/npm/v/converse-mcp-server.svg)](https://www.npmjs.com/package/converse-mcp-server)
 
-An MCP (Model Context Protocol) server that lets Claude talk to other AI models. Use it to chat with models from OpenAI, Google, Anthropic, X.AI, Mistral, DeepSeek, or OpenRouter. You can either talk to one model at a time or get multiple models to weigh in on complex decisions.
+An MCP (Model Context Protocol) server that lets Claude talk to other AI models. Use it to chat with models from OpenAI, Google, Anthropic, X.AI, Mistral, DeepSeek, OpenRouter, or Abliteration. You can either talk to one model at a time or get multiple models to weigh in on complex decisions.
 
 ## 📋 Requirements
 
@@ -25,6 +25,7 @@ You need at least one API key from these providers:
 | **Mistral**       | [console.mistral.ai](https://console.mistral.ai/)                            | `wfBMkWL0...`           |
 | **DeepSeek**      | [platform.deepseek.com](https://platform.deepseek.com/)                      | `sk-...`                |
 | **OpenRouter**    | [openrouter.ai/keys](https://openrouter.ai/keys)                             | `sk-or-...`             |
+| **Abliteration**  | [abliteration.ai/console/api-keys](https://abliteration.ai/console/api-keys) | `ak_...`                |
 | **Codex**         | ChatGPT login (system-wide)                                                  | Local agentic assistant |
 
 **Note:** Codex uses your ChatGPT login (not an API key). If you have an active ChatGPT session, Codex will work automatically. For headless/server deployments, set `CODEX_API_KEY` in your environment.
@@ -43,6 +44,7 @@ claude mcp add converse \
   -e MISTRAL_API_KEY=your_key_here \
   -e DEEPSEEK_API_KEY=your_key_here \
   -e OPENROUTER_API_KEY=your_key_here \
+  -e ABLITERATION_API_KEY=ak_your_key_here \
   -e ENABLE_RESPONSE_SUMMARIZATION=true \
   -e SUMMARIZATION_MODEL=gpt-5-nano \
   -s user \
@@ -67,6 +69,7 @@ Add this configuration to your Claude Desktop settings:
         "MISTRAL_API_KEY": "your_key_here",
         "DEEPSEEK_API_KEY": "your_key_here",
         "OPENROUTER_API_KEY": "your_key_here",
+        "ABLITERATION_API_KEY": "ak_your_key_here",
         "ENABLE_RESPONSE_SUMMARIZATION": "true",
         "SUMMARIZATION_MODEL": "gpt-5-nano"
       }
@@ -304,6 +307,16 @@ Thinking mode maps `reasoning_effort` to `none` (off), `high` (enabled levels up
 
 Any other model works via its full `provider/model` slug or the `openrouter:` namespace — no extra configuration. Append `:online` to a slug to opt into web search (adds a real per-request cost).
 
+### Abliteration Models
+
+Abliteration provides uncensored ("abliterated") reasoning models through its OpenAI-compatible Chat Completions API at `https://api.abliteration.ai/v1`; see [its documentation](https://docs.abliteration.ai). Use the `abliteration` or `ablit` namespace; set `ABLITERATION_DEFAULT_MODEL` to override the default. Web search is not supported.
+
+- **abliterated-model-large-v2** (default; aliases: `abliterated-large`, `abliterated-large-v2`): GLM-5.3-derived, text-only (1M context, 999,990 max output)
+- **abliterated-model-large** (alias: `abliterated-large-v1`): Previous large model, GLM-5.2-derived, text-only (1M context, 999,990 max output)
+- **abliterated-model** (aliases: `abliterated`, `abliterated-base`): Multimodal with image input (256K context, 262,134 max output)
+
+All models reason by default and return reasoning traces (`reasoning_content`), streamed as thinking. `reasoning_effort` is clamped to each model's supported levels: large-v2 runs `low`/`high`/`max` and cannot disable reasoning (`none`/`minimal`/`low` → `low`, `medium`/`high` → `high`, `xhigh`/`max` → `max`); large runs `high`/`max` and can disable (`none` → `none`, `minimal`–`high` → `high`, `xhigh`/`max` → `max`); base accepts every level, with `none` disabling reasoning. Only `abliterated-model` accepts images; image requests to the text-only large models are rejected before sending.
+
 ### Codex Models
 
 OpenAI Codex agentic coding assistant. `codex` uses its default model (GPT-6 Astra, or `CODEX_DEFAULT_MODEL`); `codex:<model>` picks one (e.g. `codex:luna`, `codex:astra`, `codex:gpt-5.6-terra`):
@@ -369,6 +382,7 @@ ANTHROPIC_API_KEY=sk-ant-your_anthropic_key_here
 MISTRAL_API_KEY=your_mistral_key_here
 DEEPSEEK_API_KEY=your_deepseek_key_here
 OPENROUTER_API_KEY=sk-or-your_openrouter_key_here
+ABLITERATION_API_KEY=ak_your_key_here
 
 # Optional: Server configuration
 PORT=3157
@@ -402,6 +416,7 @@ ANTHROPIC_DEFAULT_MODEL=claude-opus-5-5
 MISTRAL_DEFAULT_MODEL=mistral-medium-3-5
 DEEPSEEK_DEFAULT_MODEL=deepseek-v4-pro
 OPENROUTER_DEFAULT_MODEL=z-ai/glm-5.2        # any vendor/model slug is accepted
+ABLITERATION_DEFAULT_MODEL=abliterated-model-large-v2
 ```
 
 ### Configuration Options
@@ -506,6 +521,7 @@ Provider priority order (subscription-based local providers first, then API-key
 9. Mistral (`mistral` → Mistral Medium 3.5)
 10. DeepSeek (`deepseek` → DeepSeek V4 Pro)
 11. OpenRouter (`openrouter` → GLM 5.2)
+12. Abliteration (`abliteration` / `ablit` → `abliterated-model-large-v2`)
 
 **Local agent permissions:** Bare model names and `auto` reach the local agent providers whenever they are set up, not only when named explicitly. The Antigravity CLI runs `agy` with `--dangerously-skip-permissions` because headless calls cannot prompt for tool approval — every tool request is auto-approved, including shell commands and file writes. The Claude Agent SDK runs with `bypassPermissions`. Codex uses `CODEX_SANDBOX_MODE` (read-only by default). A read-only prompt is not an enforced security boundary for these providers, so use them only with trusted prompts and context. To keep a request on a plain API, name the provider: `google:pro`, `anthropic:opus`, `openai:gpt-6-astra`.
 
@@ -532,7 +548,8 @@ If you've cloned the repository locally:
         "ANTHROPIC_API_KEY": "your_key_here",
         "MISTRAL_API_KEY": "your_key_here",
         "DEEPSEEK_API_KEY": "your_key_here",
-        "OPENROUTER_API_KEY": "your_key_here"
+        "OPENROUTER_API_KEY": "your_key_here",
+        "ABLITERATION_API_KEY": "ak_your_key_here"
       }
     }
   }
@@ -747,6 +764,7 @@ converse/
 │   │   ├── mistral.js        # Mistral AI provider
 │   │   ├── deepseek.js       # DeepSeek provider
 │   │   ├── openrouter.js     # OpenRouter provider
+│   │   ├── abliteration.js   # Abliteration provider
 │   │   ├── openrouter-discovery.js # Request-local OpenRouter slug discovery
 │   │   ├── openai-compatible.js # Base for OpenAI-compatible APIs
 │   │   ├── codex.js          # Codex agentic SDK provider

package/docs/API.md

@@ -564,6 +564,16 @@ Models with adaptive thinking control depth via `reasoning_effort`, which is pas
 
 Any other model works via its full `provider/model` slug (e.g. `anthropic/claude-sonnet-5`) or the `openrouter:` namespace. Append `:online` to a slug (e.g. `z-ai/glm-5.2:online`) to opt into web search, which adds a real per-request cost.
 
+### Abliteration Models
+
+| Model | Aliases | Context | Output | Notes |
+|-------|---------|---------|--------|-------|
+| `abliterated-model-large-v2` | `abliterated-large`, `abliterated-large-v2` | 1M | 999,990 | Default; GLM-5.3-derived, text-only |
+| `abliterated-model-large` | `abliterated-large-v1` | 1M | 999,990 | Previous large model, GLM-5.2-derived, text-only |
+| `abliterated-model` | `abliterated`, `abliterated-base` | 256K | 262,134 | Multimodal, image input |
+
+All three are uncensored ("abliterated") reasoning models using Abliteration's OpenAI-compatible Chat Completions API at `https://api.abliteration.ai/v1`. Get an API key at [abliteration.ai/console/api-keys](https://abliteration.ai/console/api-keys); docs: [docs.abliteration.ai](https://docs.abliteration.ai). Route with the `abliteration:` namespace (canonical) or `ablit:`. All reason by default and return `reasoning_content`, streamed as thinking. `reasoning_effort` is clamped to each model's supported levels: large-v2 runs `low`/`high`/`max` and cannot disable reasoning (`none`/`minimal`/`low` → `low`, `medium`/`high` → `high`, `xhigh`/`max` → `max`); large runs `high`/`max` and can disable (`none` → `none`, `minimal`–`high` → `high`, `xhigh`/`max` → `max`); base accepts every level, with `none` disabling reasoning. Only `abliterated-model` accepts images; image requests to the text-only large models are rejected before sending. Web search is not supported.
+
 ### Codex (agentic, local)
 
 **Codex** is an agentic coding assistant with direct filesystem access:
@@ -632,9 +642,9 @@ Reach these only with the `copilot:` namespace (also `github-copilot:`, `copilot
 Every entry in `models` takes one of four forms:
 
 - **`auto`** — the first available provider's default model.
-- **`provider`** — that provider's default model (hardcoded, or `<PROVIDER>_DEFAULT_MODEL`). Namespaces: `codex`; `gemini`/`agy`/`antigravity`/`gemini-cli` (Antigravity CLI); `claude`/`claude-code`/`claude-sdk` (Claude Agent SDK); `copilot`/`github-copilot`/`copilot-sdk`; `openai`; `google`; `xai`; `anthropic`; `mistral`; `deepseek`; `openrouter`.
+- **`provider`** — that provider's default model (hardcoded, or `<PROVIDER>_DEFAULT_MODEL`). Namespaces: `codex`; `gemini`/`agy`/`antigravity`/`gemini-cli` (Antigravity CLI); `claude`/`claude-code`/`claude-sdk` (Claude Agent SDK); `copilot`/`github-copilot`/`copilot-sdk`; `openai`; `google`; `xai`; `anthropic`; `mistral`; `deepseek`; `openrouter`; `abliteration`; `ablit`.
 - **`provider:model`** — that model on that provider only; the model must be in that provider's list.
-- **Bare `model`** — the first provider, in the order `codex`, `gemini-cli`, `claude`, `openai`, `google`, `xai`, `anthropic`, `mistral`, `deepseek`, `openrouter`, whose list contains the ID or alias and that is set up (API key for API providers; SDK plus login for Codex and the Claude Agent SDK; the `agy` binary for Antigravity). On an authentication or availability error, the next set-up provider serving **the same model** takes over; a provider whose alias points at a different model is never substituted (bare `fable` is Fable 5.1 on the Claude Agent SDK and Fable 5 on the Anthropic API). Copilot never serves bare names.
+- **Bare `model`** — the first provider, in the order `codex`, `gemini-cli`, `claude`, `openai`, `google`, `xai`, `anthropic`, `mistral`, `deepseek`, `openrouter`, `abliteration`, whose list contains the ID or alias and that is set up (API key for API providers; SDK plus login for Codex and the Claude Agent SDK; the `agy` binary for Antigravity). On an authentication or availability error, the next set-up provider serving **the same model** takes over; a provider whose alias points at a different model is never substituted (bare `fable` is Fable 5.1 on the Claude Agent SDK and Fable 5 on the Anthropic API). Copilot never serves bare names.
 
 **Unknown names are rejected**, never forwarded: the error lists up to three close matches, e.g. `Unknown model "gtp-6-astra". Did you mean: gpt-6-astra?` or `Unknown openai model "spark" in "openai:spark". Did you mean: codex:spark?`. The exception is OpenRouter: a full `vendor/model` slug (bare or `openrouter:`) is validated against OpenRouter's live catalog.
 
@@ -650,6 +660,7 @@ Every entry in `models` takes one of four forms:
 "mistral"                  // Mistral (-> mistral-medium-3-5)
 "z-ai/glm-5.2"             // OpenRouter (full slug)
 "z-ai/glm-5.2:online"      // OpenRouter with web search opt-in
+"ablit"                    // Abliteration (-> abliterated-model-large-v2)
 "fable"                    // Claude Agent SDK (-> claude-fable-5-1) when set up, otherwise Anthropic API (-> claude-fable-5)
 "opus"                     // Claude Agent SDK (-> claude-opus-5-5), else Anthropic API
 "anthropic:opus"           // Anthropic API only
@@ -666,7 +677,7 @@ Every entry in `models` takes one of four forms:
 - **chat mode**: `["auto"]` selects the first available provider and uses its default model, with failover to the next provider on error.
 - **consensus mode**: `["auto"]` expands to the first 3 available providers.
 
-Provider auto-selection priority (subscription-based CLI/SDK providers first, then API-key providers): `codex`, `gemini-cli`, `claude`, `copilot`, `openai`, `google`, `xai`, `anthropic`, `mistral`, `deepseek`, `openrouter`.
+Provider auto-selection priority (subscription-based CLI/SDK providers first, then API-key providers): `codex`, `gemini-cli`, `claude`, `copilot`, `openai`, `google`, `xai`, `anthropic`, `mistral`, `deepseek`, `openrouter`, `abliteration`.
 
 ## Configuration
 
@@ -714,6 +725,7 @@ ANTHROPIC_DEFAULT_MODEL=claude-opus-5-5
 MISTRAL_DEFAULT_MODEL=mistral-medium-3-5
 DEEPSEEK_DEFAULT_MODEL=deepseek-v4-pro
 OPENROUTER_DEFAULT_MODEL=z-ai/glm-5.2        # any vendor/model slug is accepted
+ABLITERATION_DEFAULT_MODEL=abliterated-model-large-v2
 ```
 
 ## Context Processing
@@ -822,6 +834,7 @@ ANTHROPIC_API_KEY=sk-ant-...
 MISTRAL_API_KEY=...
 DEEPSEEK_API_KEY=...
 OPENROUTER_API_KEY=sk-or-...
+ABLITERATION_API_KEY=ak_...
 TYPESAFE_API_KEY=...            # decide tool only
 ```
 

package/docs/PROVIDERS.md

@@ -99,6 +99,17 @@ This guide documents all supported AI providers in the Converse MCP Server and t
 - **Reasoning**: Per-model. `z-ai/glm-5.2` and the `deepseek/deepseek-v4-*` slugs are effort-tiered (`max` → `xhigh`, other enabled levels → `high`, `none` disables); `qwen/qwen3.7-*` and `moonshotai/kimi-k2.6` are enable/disable only (`none` disables, any other level enables); `moonshotai/kimi-k2.7-code` always reasons and cannot be disabled; `openrouter/auto` lets the router choose.
 - **Web search (opt-in, adds cost)**: OpenRouter web search is off by default because it incurs a real per-request charge. Enable it explicitly by appending `:online` to a slug (e.g. `z-ai/glm-5.2:online` or `openrouter:qwen/qwen3.7-max:online`). When enabled, `annotations[].url_citation` citations are captured into metadata. Ordinary requests never attach a web-search plugin.
 
+### Abliteration
+- **API Key Format**: `ak_...`
+- **Get Key**: [abliteration.ai/console/api-keys](https://abliteration.ai/console/api-keys) ([documentation](https://docs.abliteration.ai))
+- **Environment Variable**: `ABLITERATION_API_KEY`
+- **API**: OpenAI-compatible Chat Completions API at `https://api.abliteration.ai/v1`
+- **Supported Models**:
+  - `abliterated-model-large-v2` (default; aliases: `abliterated-large`, `abliterated-large-v2`) - Uncensored, GLM-5.3-derived reasoning model, text-only (1M context, 999,990 max output)
+  - `abliterated-model-large` (alias: `abliterated-large-v1`) - Previous large model, GLM-5.2-derived, text-only (1M context, 999,990 max output)
+  - `abliterated-model` (aliases: `abliterated`, `abliterated-base`) - Multimodal with image input (256K context, 262,134 max output)
+- **Reasoning**: All models reason by default and return `reasoning_content`, streamed as thinking. The large-v2 model supports `low`/`high`/`max` and cannot disable reasoning (`none`/`minimal`/`low` → `low`, `medium`/`high` → `high`, `xhigh`/`max` → `max`). The large model supports `high`/`max` and can disable reasoning (`none` → `none`, `minimal`–`high` → `high`, `xhigh`/`max` → `max`). The base model accepts every level, with `none` disabling reasoning.
+
 ### Codex
 - **API Key Format**: Optional (uses ChatGPT login by default)
 - **Authentication**: ChatGPT login (system-wide) OR `CODEX_API_KEY`
@@ -265,6 +276,7 @@ DEEPSEEK_API_KEY=your_deepseek_key_here
 
 # OpenRouter needs only an API key; any provider/model slug works directly
 OPENROUTER_API_KEY=sk-or-your_key_here
+ABLITERATION_API_KEY=ak_your_key_here
 # Optional: referer and title for OpenRouter ranking credit
 OPENROUTER_REFERER=https://github.com/YourUsername/YourApp
 OPENROUTER_TITLE=Converse
@@ -292,6 +304,7 @@ ANTHROPIC_DEFAULT_MODEL=claude-opus-5-5
 MISTRAL_DEFAULT_MODEL=mistral-medium-3-5
 DEEPSEEK_DEFAULT_MODEL=deepseek-v4-pro
 OPENROUTER_DEFAULT_MODEL=z-ai/glm-5.2        # any vendor/model slug is accepted
+ABLITERATION_DEFAULT_MODEL=abliterated-model-large-v2
 ```
 
 ### Claude Configuration (claude_desktop_config.json)
@@ -307,6 +320,7 @@ OPENROUTER_DEFAULT_MODEL=z-ai/glm-5.2        # any vendor/model slug is accepted
         "MISTRAL_API_KEY": "your_key_here",
         "DEEPSEEK_API_KEY": "your_key_here",
         "OPENROUTER_API_KEY": "your_key_here",
+        "ABLITERATION_API_KEY": "ak_your_key_here",
         "OPENROUTER_REFERER": "https://github.com/YourUsername/YourApp",
         "OPENROUTER_TITLE": "Converse"
       }
@@ -324,12 +338,13 @@ All providers support streaming responses for real-time output.
 - **Full Support**: OpenAI, Google, X.AI (Grok 4.5), Anthropic (Claude-4 series, Claude-3-Opus)
 - **Mistral**: `mistral-medium-3-5` and `mistral-small-2603` accept images; `mistral-large-2512` is text-only
 - **Via OpenRouter**: Depends on the model — `qwen/qwen3.7-plus`, `moonshotai/kimi-k2.7-code`, and `moonshotai/kimi-k2.6` accept images; `z-ai/glm-5.2` and the `deepseek/deepseek-v4-*` slugs are text-only
+- **Abliteration**: Only `abliterated-model` accepts images; both large models are text-only, and image requests to them are rejected before sending
 - **No Support**: DeepSeek (native), Codex
 
 ### Web Search
 - **Automatic where supported**: OpenAI, Google, and X.AI (Grok 4.5, via Agent Tools) attach web search on every request for capable models; the model decides whether to use it
 - **OpenRouter (opt-in, adds cost)**: Off by default; append `:online` to a slug to enable it per request (real per-request charge), with citations captured into metadata
-- **No Support**: Anthropic, Mistral, DeepSeek, Codex
+- **No Support**: Anthropic, Mistral, DeepSeek, Codex, Abliteration
 
 ### Thinking/Reasoning Modes
 - **OpenAI**: GPT-5 family and O3 series models support the `reasoning_effort` parameter (GPT-5.6 accepts `none` through `max`, mapping `minimal` to `low`; GPT-5 Pro is fixed at `high`)
@@ -341,6 +356,7 @@ All providers support streaming responses for real-time output.
 - **Mistral**: `mistral-medium-3-5` and `mistral-small-2603` map `reasoning_effort` to `high` (enabled) or `none` (disabled); `mistral-large-2512` has no adjustable reasoning
 - **DeepSeek**: V4 models use thinking mode via `reasoning_effort` (`none` disables; enabled levels use `high`, `max` uses `max`)
 - **OpenRouter**: Reasoning is per-model (effort-tiered, enable/disable-only, mandatory, or router-chosen — see the OpenRouter section)
+- **Abliteration**: All models reason by default and return reasoning traces (`reasoning_content`), streamed as thinking; see the Abliteration section for `reasoning_effort` clamping
 - **Codex**: Thread-based agentic reasoning with persistent context
 - **Others**: Standard inference only
 
@@ -362,11 +378,11 @@ Routing is derived entirely from each provider's model list (canonical IDs plus
    - `gemini`, `agy`, `antigravity`, `gemini-cli` → Gemini via Antigravity CLI
    - `claude`, `claude-code`, `claude-sdk` → Claude Agent SDK
    - `copilot`, `github-copilot`, `copilot-sdk` → Copilot SDK
-   - `openai`, `google`, `xai`, `anthropic`, `mistral`, `deepseek`, `openrouter` → the matching API provider
+   - `openai`, `google`, `xai`, `anthropic`, `mistral`, `deepseek`, `openrouter`, `abliteration`, `ablit` → the matching API provider
 
 2. **`provider:model`** — that model on that provider only (e.g. `codex:astra`, `gemini:pro`, `google:pro`, `anthropic:opus`, `copilot:sonnet`). The model must be in that provider's list; there is no failover to another provider.
 
-3. **Bare `model`** — an ID or alias without a namespace goes to the first provider, in this order, whose list contains the name and that is set up: Codex, Antigravity CLI, Claude Agent SDK, OpenAI, Google, X.AI, Anthropic, Mistral, DeepSeek, OpenRouter.
+3. **Bare `model`** — an ID or alias without a namespace goes to the first provider, in this order, whose list contains the name and that is set up: Codex, Antigravity CLI, Claude Agent SDK, OpenAI, Google, X.AI, Anthropic, Mistral, DeepSeek, OpenRouter, Abliteration.
    - "Set up" means an API key for API providers; for Codex, the SDK plus a login file or `CODEX_API_KEY`; for the Claude Agent SDK, the SDK plus a login file, a macOS login, or `CLAUDE_CODE_OAUTH_TOKEN`/`ANTHROPIC_API_KEY`; for Antigravity, the `agy` binary.
    - If that provider fails with an authentication or availability error (including an expired login, which is only detected at call time), the next set-up provider that serves **the same model** takes over. A provider whose alias of that name points at a different model is never substituted (bare `fable` is Fable 5.1 on the Claude Agent SDK and Fable 5 on the Anthropic API; bare `flash` is Gemini 3.8 Flash on Antigravity and Gemini 2.5 Flash on the Google API).
    - Copilot never serves bare names; use `copilot:<model>`.
@@ -400,6 +416,7 @@ Examples:
 "z-ai/glm-5.2:online"      // OpenRouter with web search opt-in
 "anthropic/claude-sonnet-5" // OpenRouter (any full slug routes as-is)
 "openrouter/auto"          // OpenRouter auto-selection
+"ablit"                    // Abliteration default model (abliterated-model-large-v2)
 ```
 
 ```json

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "converse-mcp-server",
-  "version": "4.4.0",
+  "version": "4.5.0",
   "description": "Converse MCP Server - Converse with other LLMs with chat and consensus tools",
   "type": "module",
   "main": "src/index.js",
@@ -93,20 +93,20 @@
     ".env.example"
   ],
   "dependencies": {
-    "@anthropic-ai/claude-agent-sdk": "^0.3.284",
+    "@anthropic-ai/claude-agent-sdk": "^0.3.285",
     "@anthropic-ai/sdk": "^0.129.0",
     "@github/copilot-sdk": "^1.0.15",
     "@google/genai": "^2.24.0",
     "@lydell/node-pty": "1.2.0-beta.15",
     "@mistralai/mistralai": "^2.7.0",
     "@modelcontextprotocol/sdk": "^1.31.0",
-    "@openai/codex-sdk": "^0.159.0",
+    "@openai/codex-sdk": "^0.159.2",
     "cors": "^2.8.6",
     "dotenv": "^18.0.4",
     "express": "^5.2.1",
     "lru-cache": "^11.5.3",
     "nanoid": "^6.0.1",
-    "openai": "^7.23.0",
+    "openai": "^7.25.0",
     "p-limit": "^7.3.3",
     "vite": "^8.3.1"
   },

package/src/async/providerStreamNormalizer.js

@@ -9,7 +9,7 @@
  * - error: Error event with recovery information
  *
  * This normalizer enables seamless provider switching and uniform async processing
- * across all supported LLM providers (OpenAI, Google, XAI, Anthropic, Mistral, DeepSeek, OpenRouter).
+ * across all supported LLM providers (OpenAI, Google, XAI, Anthropic, Mistral, DeepSeek, OpenRouter, Abliteration).
  */
 
 import { debugLog, debugError } from '../utils/console.js';
@@ -40,8 +40,11 @@ class ProviderStreamNormalizer {
       google: this.normalizeGoogleStream.bind(this),
       anthropic: this.normalizeAnthropicStream.bind(this),
       mistral: this.normalizeMistralStream.bind(this),
-      deepseek: this.normalizeDeepSeekStream.bind(this),
+      deepseek: (stream, context) =>
+        this.normalizeChatCompletionsStream(stream, context, 'deepseek'),
       openrouter: this.normalizeOpenRouterStream.bind(this),
+      abliteration: (stream, context) =>
+        this.normalizeChatCompletionsStream(stream, context, 'abliteration'),
       codex: this.normalizeCodexStream.bind(this),
       copilot: this.normalizePassthroughStream.bind(this),
       claude: this.normalizePassthroughStream.bind(this),
@@ -507,10 +510,11 @@ class ProviderStreamNormalizer {
   }
 
   /**
-   * Normalize DeepSeek streaming format
+   * Normalize the event stream of a provider built on the shared
+   * OpenAI-compatible Chat Completions base (DeepSeek, Abliteration), where
+   * streamed `reasoning_content` arrives as separate thinking events.
    */
-  async *normalizeDeepSeekStream(stream, context) {
-    const provider = 'deepseek';
+  async *normalizeChatCompletionsStream(stream, context, provider) {
     const model = context.model || 'unknown';
     const startTime = Date.now();
 
@@ -530,7 +534,7 @@ class ProviderStreamNormalizer {
           continue;
         }
 
-        // Handle delta events (including reasoning tokens for DeepSeek-R1)
+        // Handle delta events
         if (event.type === 'delta') {
           accumulatedContent += event.content || '';
           yield this.createDeltaEvent(event.content || '', provider, model, {
@@ -546,7 +550,7 @@ class ProviderStreamNormalizer {
           continue;
         }
 
-        // Handle usage events (with DeepSeek-specific reasoning tokens)
+        // Handle usage events (with reasoning token counts when reported)
         if (event.type === 'usage') {
           accumulatedUsage = event.usage;
           if (event.usage.reasoning_tokens) {
@@ -590,7 +594,7 @@ class ProviderStreamNormalizer {
         }
       }
     } catch (error) {
-      debugError('[StreamNormalizer] DeepSeek stream error:', error);
+      debugError(`[StreamNormalizer] ${provider} stream error:`, error);
       yield this.createErrorEvent(error, provider);
       throw error;
     }

package/src/config.js

@@ -214,6 +214,12 @@ const CONFIG_SCHEMA = {
       secret: true,
       description: 'OpenRouter API key',
     },
+    ABLITERATION_API_KEY: {
+      type: 'string',
+      required: false,
+      secret: true,
+      description: 'abliteration.ai API key',
+    },
     TYPESAFE_API_KEY: {
       type: 'string',
       required: false,
@@ -438,6 +444,8 @@ function validateApiKeyFormat(provider, apiKey) {
     return apiKey.length >= 32; // DeepSeek keys are typically 32+ chars
   case 'openrouter':
     return apiKey.startsWith('sk-or-') && apiKey.length >= 40;
+  case 'abliteration':
+    return apiKey.startsWith('ak_') && apiKey.length > 10;
   default:
     return apiKey.length >= 10; // Basic minimum length check
   }
@@ -713,7 +721,7 @@ export async function loadConfig() {
 
     if (availableKeys.length === 0 && !hasVertexAI && !hasSdkProvider) {
       errors.push(
-        'At least one API key must be configured: OPENAI_API_KEY, XAI_API_KEY, GOOGLE_API_KEY, GEMINI_API_KEY, ANTHROPIC_API_KEY, MISTRAL_API_KEY, DEEPSEEK_API_KEY, OPENROUTER_API_KEY, or TYPESAFE_API_KEY. Alternatively, configure Google Vertex AI or use an SDK-based provider (codex, claude, copilot) or the Antigravity CLI (gemini-cli).',
+        'At least one API key must be configured: OPENAI_API_KEY, XAI_API_KEY, GOOGLE_API_KEY, GEMINI_API_KEY, ANTHROPIC_API_KEY, MISTRAL_API_KEY, DEEPSEEK_API_KEY, OPENROUTER_API_KEY, ABLITERATION_API_KEY, or TYPESAFE_API_KEY. Alternatively, configure Google Vertex AI or use an SDK-based provider (codex, claude, copilot) or the Antigravity CLI (gemini-cli).',
       );
     }
 

package/src/prompts/helpPrompt.js

@@ -374,6 +374,7 @@ export function generateHelpContent(config = null) {
     mistral: providers.mistral?.getSupportedModels() || {},
     deepseek: providers.deepseek?.getSupportedModels() || {},
     openrouter: providers.openrouter?.getSupportedModels() || {},
+    abliteration: providers.abliteration?.getSupportedModels() || {},
     // CLI providers - use safeGetModels (may throw if CLI not installed)
     codex: safeGetModels(providers.codex, 'codex'),
     claude: safeGetModels(providers.claude, 'claude'),
@@ -487,6 +488,7 @@ ${formatProviderModels('Anthropic', allModels.anthropic)}
 ${formatProviderModels('Mistral', allModels.mistral)}
 ${formatProviderModels('DeepSeek', allModels.deepseek)}
 ${formatProviderModels('OpenRouter', allModels.openrouter)}
+${formatProviderModels('Abliteration', allModels.abliteration)}
 ${formatProviderModels('Codex', allModels.codex)}
 ${formatProviderModels('Claude CLI', allModels.claude)}
 ${formatProviderModels('Gemini (Antigravity CLI)', allModels['gemini-cli'])}

package/src/providers/abliteration.js

@@ -0,0 +1,119 @@
+/**
+ * Abliteration Provider
+ *
+ * Provider implementation for abliteration.ai models using its OpenAI-compatible
+ * Chat Completions API.
+ * Implements the unified interface: async invoke(messages, options) => { content, stop_reason, rawResponse }
+ *
+ * Chat Completions is used rather than the Responses surface: abliteration.ai's
+ * /v1/responses is stateless (no previous_response_id to gain from), rejects
+ * `max` effort on the base model, while Chat Completions accepts every effort
+ * tier on all models and returns the trace as `reasoning_content`, which the
+ * shared OpenAI-compatible base already surfaces.
+ */
+
+import { createOpenAICompatibleProvider } from './openai-compatible.js';
+import { debugLog } from '../utils/console.js';
+import {
+  clampReasoningEffort,
+  EFFORT_LADDER,
+} from '../utils/reasoningEffort.js';
+
+// All three models always reason by default. `reasoningTiers` lists the
+// distinct depths each model actually runs; the API accepts the whole ladder
+// but silently collapses it onto these, so clamping here keeps the requested
+// tier and the tier that runs in agreement.
+const SUPPORTED_MODELS = {
+  'abliterated-model-large-v2': {
+    modelName: 'abliterated-model-large-v2',
+    friendlyName: 'Abliterated Large V2',
+    contextWindow: 1000000,
+    maxOutputTokens: 999990,
+    supportsStreaming: true,
+    supportsImages: false, // Text-only; image content is rejected with a 400
+    supportsWebSearch: false,
+    supportsReasoning: true,
+    // Cannot disable reasoning: `none` would run at low with the trace hidden,
+    // so it clamps up to an explicit `low` that keeps the trace visible.
+    reasoningTiers: ['low', 'high', 'max'],
+    supportsJsonOutput: true,
+    supportsFunctionCalling: true,
+    timeout: 1800000,
+    description:
+      'Abliterated Large V2 - uncensored GLM-5.3 derived reasoning model with 1M context',
+    aliases: ['abliterated-large', 'abliterated-large-v2'],
+  },
+  'abliterated-model-large': {
+    modelName: 'abliterated-model-large',
+    friendlyName: 'Abliterated Large',
+    contextWindow: 1000000,
+    maxOutputTokens: 999990,
+    supportsStreaming: true,
+    supportsImages: false, // Text-only; image content is rejected with a 400
+    supportsWebSearch: false,
+    supportsReasoning: true,
+    reasoningTiers: ['none', 'high', 'max'],
+    supportsJsonOutput: true,
+    supportsFunctionCalling: true,
+    timeout: 1800000,
+    description:
+      'Abliterated Large - previous uncensored GLM-5.2 derived reasoning model with 1M context',
+    aliases: ['abliterated-large-v1'],
+  },
+  'abliterated-model': {
+    modelName: 'abliterated-model',
+    friendlyName: 'Abliterated Model',
+    contextWindow: 262144,
+    maxOutputTokens: 262134,
+    supportsStreaming: true,
+    supportsImages: true,
+    supportsWebSearch: false,
+    supportsReasoning: true,
+    reasoningTiers: EFFORT_LADDER,
+    supportsJsonOutput: true,
+    supportsFunctionCalling: true,
+    timeout: 900000,
+    description:
+      'Abliterated Model - uncensored multimodal reasoning model with 256K context',
+    aliases: ['abliterated', 'abliterated-base'],
+  },
+};
+
+/**
+ * abliteration.ai API keys are issued as `ak_...` tokens.
+ */
+function validateApiKey(apiKey) {
+  if (!apiKey || typeof apiKey !== 'string') {
+    return false;
+  }
+
+  return apiKey.startsWith('ak_') && apiKey.length > 10;
+}
+
+/**
+ * Attach `reasoning_effort` clamped onto the tiers the resolved model runs.
+ * Capability-gated: an unknown pass-through ID never receives the field and
+ * runs at the server's default depth.
+ */
+async function transformRequest(requestPayload, context = {}) {
+  const { modelConfig, reasoningEffort } = context;
+
+  if (modelConfig?.supportsReasoning && reasoningEffort) {
+    requestPayload.reasoning_effort = clampReasoningEffort(
+      reasoningEffort,
+      modelConfig.reasoningTiers,
+    );
+  }
+
+  debugLog('[Abliteration] Request payload prepared');
+  return requestPayload;
+}
+
+export const abliterationProvider = createOpenAICompatibleProvider({
+  baseURL: 'https://api.abliteration.ai/v1',
+  providerName: 'Abliteration',
+  supportedModels: SUPPORTED_MODELS,
+  defaultModel: 'abliterated-model-large-v2',
+  validateApiKey,
+  transformRequest,
+});

package/src/providers/index.js

@@ -17,6 +17,7 @@ import { codexProvider } from './codex.js';
 import { geminiCliProvider } from './gemini-cli.js';
 import { claudeProvider } from './claude.js';
 import { copilotProvider } from './copilot.js';
+import { abliterationProvider } from './abliteration.js';
 
 /**
  * Provider registry map
@@ -34,6 +35,7 @@ const providers = {
   mistral: mistralProvider,
   deepseek: deepseekProvider,
   openrouter: openrouterProvider,
+  abliteration: abliterationProvider,
   codex: codexProvider,
   claude: claudeProvider,
   copilot: copilotProvider,

package/src/tools/chat.js

@@ -1165,7 +1165,7 @@ chatTool.inputSchema = {
       items: { type: 'string' },
       minItems: 1,
       description:
-        'Models to use. Examples: ["auto"] (recommended), ["codex"], ["codex", "gemini", "claude"], ["codex:astra"], ["gpt-6-astra"]. Forms: "provider" (its default model), "provider:model" (that provider only), or a bare "model" (served by the first configured provider that offers it, local CLI providers first, failing over to the next). Providers: codex, gemini (agy), claude, copilot, openai, google, xai, anthropic, mistral, deepseek, openrouter. Unknown names are rejected with suggestions. In mode "chat" each model answers independently; in "consensus" they refine after seeing each other; in "roundtable" they speak in the given ORDER, each seeing the transcript. Default: ["auto"].',
+        'Models to use. Examples: ["auto"] (recommended), ["codex"], ["codex", "gemini", "claude"], ["codex:astra"], ["gpt-6-astra"]. Forms: "provider" (its default model), "provider:model" (that provider only), or a bare "model" (served by the first configured provider that offers it, local CLI providers first, failing over to the next). Providers: codex, gemini (agy), claude, copilot, openai, google, xai, anthropic, mistral, deepseek, openrouter, abliteration. Unknown names are rejected with suggestions. In mode "chat" each model answers independently; in "consensus" they refine after seeing each other; in "roundtable" they speak in the given ORDER, each seeing the transcript. Default: ["auto"].',
     },
     mode: {
       type: 'string',

package/src/utils/modelRouting.js

@@ -40,6 +40,7 @@ export const PROVIDER_PRIORITY = [
   'mistral',
   'deepseek',
   'openrouter',
+  'abliteration',
 ];
 
 /**
@@ -70,6 +71,7 @@ export const PROVIDER_NAMESPACES = {
   mistral: ['mistral'],
   deepseek: ['deepseek'],
   openrouter: ['openrouter'],
+  abliteration: ['abliteration', 'ablit'],
 };
 
 /**
@@ -89,6 +91,7 @@ export const DEFAULT_MODEL_ENV_VARS = {
   mistral: 'MISTRAL_DEFAULT_MODEL',
   deepseek: 'DEEPSEEK_DEFAULT_MODEL',
   openrouter: 'OPENROUTER_DEFAULT_MODEL',
+  abliteration: 'ABLITERATION_DEFAULT_MODEL',
 };
 
 /**
@@ -583,6 +586,7 @@ const API_KEY_ENV_VARS = {
   mistral: 'MISTRAL_API_KEY',
   deepseek: 'DEEPSEEK_API_KEY',
   openrouter: 'OPENROUTER_API_KEY',
+  abliteration: 'ABLITERATION_API_KEY',
 };
 
 const LOCAL_PROVIDER_SETUP_HINTS = {