feat(local-models): first-class MLX providers + 1143x local-server ports

packages/create/README.md

@@ -101,7 +101,7 @@ with the "unsigned" color theme and OSC 133 shell integration.
 
 Model gateway config presets for all four agent harnesses, pointing at the
 unsigned model gateway (`llm.unsigned.gg/v1`) and local endpoints
-(ollama `:11434`, llama.cpp `:8080`, vLLM `:8000`). Copy-if-absent — never
+(ollama `:11434`, llama.cpp `:11435`, vLLM `:11438`). Copy-if-absent — never
 clobbers live config. Requires `UNSIGNED_LLM_API_KEY` for gateway access.
 
 ### `skills` — agentskills.io symlink installer

packages/harness-hermes/README.md

@@ -24,7 +24,7 @@ the upstream script resolves to, without executing unreviewed remote code.
 
 - **`config/cli-config.local-endpoint.yaml`** → seeded to
   `~/.hermes/cli-config.yaml` if absent: points hermes at a local
-  OpenAI-compatible endpoint (vLLM :8000 default; ollama/llama.cpp are a
+  OpenAI-compatible endpoint (vLLM :11438 default; ollama/llama.cpp are a
   one-line `openai_base_url` change). **Config-key caveat:** validate keys
   against `hermes config list` on your installed version — upstream moves
   fast (0.18.0 "The Judgment Release" is 2026-07-01); the preset is

packages/harness-hermes/config/cli-config.local-endpoint.yaml

@@ -88,9 +88,36 @@ providers:
       - qwen3.6-35b-fast
       - yutori/n1.5
 
+  # mlx — Apple-silicon local (mlx_lm.server). First-class local provider on
+  # :11436 (off llama.cpp's :11435 so both co-tenant). Serve: `mlx-serve <preset>`.
+  mlx:
+    name: MLX (mlx_lm.server, local)
+    api: http://localhost:11436/v1
+    api_key: none
+    default_model: mlx-community/gpt-oss-120b-MXFP4-Q8
+    models:
+      - mlx-community/gpt-oss-120b-MXFP4-Q8
+      - mlx-community/Qwen3.6-35B-A3B-8bit
+      - mlx-community/Devstral-Small-2-24B-Instruct-2512-4bit
+      - mlx-community/Qwen3-Coder-30B-A3B-Instruct-8bit
+      - mlx-community/gemma-4-31B-it-OptiQ-4bit
+      - mlx-community/Qwen3-VL-32B-Instruct-8bit
+
+  # rapid-mlx — alternate OpenAI-compatible MLX server (brew: rapid-mlx) on :11437.
+  # Pending install/eval; staged so the endpoint resolves once served.
+  rapid-mlx:
+    name: rapid-mlx (local)
+    api: http://localhost:11437/v1
+    api_key: none
+    default_model: mlx-community/gpt-oss-120b-MXFP4-Q8
+    models:
+      - mlx-community/gpt-oss-120b-MXFP4-Q8
+
 # Alternate local endpoints (swap `api`):
-#   llama.cpp: http://localhost:8080/v1   (serve/serve-llamacpp.sh)
-#   vLLM:      http://localhost:8000/v1   (serve/serve-vllm.sh)
+#   mlx (mlx_lm.server): http://localhost:11436/v1   (mlx-serve <preset>)
+#   rapid-mlx:           http://localhost:11437/v1   (brew: rapid-mlx)
+#   llama.cpp:           http://localhost:11435/v1   (serve/serve-llamacpp.sh)
+#   vLLM:                http://localhost:11438/v1   (serve/serve-vllm.sh)
 
 # Memory/skill-loop boundary vs Engram (OPS-480, 2026-07-08; see README Verdict
 # section for the full reasoning). hermes's built-in memory store (two flat

packages/harness-omp/README.md

@@ -37,7 +37,7 @@ are maintained separately.
   the same endpoints as harness-pi's `models.json` (the `llm` unsigned model
   gateway at llm.unsigned.gg — Crusoe/OpenRouter/Anthropic/MorphLLM upstreams,
   key via `UNSIGNED_LLM_API_KEY`, 1P item `unsigned-llm` — plus ollama :11434,
-  llama.cpp :8080, vLLM :8000) in omp's provider schema. **Rename note
+  llama.cpp :11435, vLLM :11438) in omp's provider schema. **Rename note
   (2026-07-07):** the gateway provider was `neuralwatt`; presets are
   copy-if-absent, so an existing `~/.omp/agent/models.yml` keeps the old
   name — hand-merge the `llm` block (the old hostname keeps working through

packages/harness-omp/config/models.yml

@@ -160,16 +160,47 @@ providers:
         contextWindow: 65536
         reasoning: true
   llamacpp:
-    baseUrl: http://localhost:8080/v1
+    baseUrl: http://localhost:11435/v1
     api: openai-completions
     auth: none
     discovery:
       type: llama.cpp
     models:
       - id: local
   vllm:
-    baseUrl: http://localhost:8000/v1
+    baseUrl: http://localhost:11438/v1
     api: openai-completions
     auth: none
     models:
       - id: served-model
+  mlx:
+    # Apple-silicon local (mlx_lm.server), first-class provider on :11436 —
+    # off llama.cpp's :11435 to co-tenant. Serve via `mlx-serve <preset>`.
+    baseUrl: http://localhost:11436/v1
+    api: openai-completions
+    auth: none
+    models:
+      - id: mlx-community/gpt-oss-120b-MXFP4-Q8
+        contextWindow: 131072
+        reasoning: true
+      - id: mlx-community/Qwen3.6-35B-A3B-8bit
+        contextWindow: 131072
+        reasoning: true
+      - id: mlx-community/Devstral-Small-2-24B-Instruct-2512-4bit
+        contextWindow: 131072
+      - id: mlx-community/Qwen3-Coder-30B-A3B-Instruct-8bit
+        contextWindow: 262144
+      - id: mlx-community/gemma-4-31B-it-OptiQ-4bit
+        contextWindow: 131072
+      - id: mlx-community/Qwen3-VL-32B-Instruct-8bit
+        contextWindow: 131072
+  rapid-mlx:
+    # Alternate OpenAI-compatible MLX server (brew: rapid-mlx) on :11437.
+    # Pending install/eval; staged so the endpoint resolves once served.
+    baseUrl: http://localhost:11437/v1
+    api: openai-completions
+    auth: none
+    models:
+      - id: mlx-community/gpt-oss-120b-MXFP4-Q8
+        contextWindow: 131072
+        reasoning: true

packages/harness-opencode/config/opencode.json

@@ -163,7 +163,7 @@
       "npm": "@ai-sdk/openai-compatible",
       "name": "llama.cpp server (local)",
       "options": {
-        "baseURL": "http://localhost:8080/v1"
+        "baseURL": "http://localhost:11435/v1"
       },
       "models": {
         "local": {
@@ -175,13 +175,38 @@
       "npm": "@ai-sdk/openai-compatible",
       "name": "vLLM (local)",
       "options": {
-        "baseURL": "http://localhost:8000/v1"
+        "baseURL": "http://localhost:11438/v1"
       },
       "models": {
         "served-model": {
           "name": "vLLM served model"
         }
       }
+    },
+    "mlx": {
+      "npm": "@ai-sdk/openai-compatible",
+      "name": "MLX (mlx_lm.server, local :11436)",
+      "options": {
+        "baseURL": "http://localhost:11436/v1"
+      },
+      "models": {
+        "mlx-community/gpt-oss-120b-MXFP4-Q8": { "name": "gpt-oss-120b (MLX flagship)" },
+        "mlx-community/Qwen3.6-35B-A3B-8bit": { "name": "Qwen3.6-35B-A3B (MLX fast)" },
+        "mlx-community/Devstral-Small-2-24B-Instruct-2512-4bit": { "name": "Devstral-2 24B (MLX code)" },
+        "mlx-community/Qwen3-Coder-30B-A3B-Instruct-8bit": { "name": "Qwen3-Coder-30B (MLX)" },
+        "mlx-community/gemma-4-31B-it-OptiQ-4bit": { "name": "Gemma-4 31B (MLX vision)" },
+        "mlx-community/Qwen3-VL-32B-Instruct-8bit": { "name": "Qwen3-VL 32B (MLX)" }
+      }
+    },
+    "rapid-mlx": {
+      "npm": "@ai-sdk/openai-compatible",
+      "name": "rapid-mlx (local :11437)",
+      "options": {
+        "baseURL": "http://localhost:11437/v1"
+      },
+      "models": {
+        "mlx-community/gpt-oss-120b-MXFP4-Q8": { "name": "gpt-oss-120b (rapid-mlx)" }
+      }
     }
   }
 }

packages/harness-pi/README.md

@@ -17,7 +17,7 @@ consumes: local-models/README.md (endpoint wiring)
 
 - **`config/models.json`** → `~/.pi/agent/models.json` (only if absent — live
   config is never clobbered): three OpenAI-compatible local providers —
-  `ollama` (:11434), `llamacpp` (:8080), `vllm` (:8000). Model ids are
+  `ollama` (:11434), `llamacpp` (:11435), `vllm` (:11438). Model ids are
   placeholders; align them with what you actually serve
   (`packages/local-models/MODELS.md`).
 - **`extensions/local-model-switch.ts`** → symlinked into

packages/harness-pi/config/models.json

@@ -224,7 +224,7 @@
       ]
     },
     "llamacpp": {
-      "baseUrl": "http://localhost:8080/v1",
+      "baseUrl": "http://localhost:11435/v1",
       "api": "openai-completions",
       "apiKey": "none",
       "models": [
@@ -234,14 +234,35 @@
       ]
     },
     "vllm": {
-      "baseUrl": "http://localhost:8000/v1",
+      "baseUrl": "http://localhost:11438/v1",
       "api": "openai-completions",
       "apiKey": "none",
       "models": [
         {
           "id": "served-model"
         }
       ]
+    },
+    "mlx": {
+      "baseUrl": "http://localhost:11436/v1",
+      "api": "openai-completions",
+      "apiKey": "none",
+      "models": [
+        { "id": "mlx-community/gpt-oss-120b-MXFP4-Q8", "contextWindow": 131072, "reasoning": true },
+        { "id": "mlx-community/Qwen3.6-35B-A3B-8bit", "contextWindow": 131072, "reasoning": true },
+        { "id": "mlx-community/Devstral-Small-2-24B-Instruct-2512-4bit", "contextWindow": 131072 },
+        { "id": "mlx-community/Qwen3-Coder-30B-A3B-Instruct-8bit", "contextWindow": 262144 },
+        { "id": "mlx-community/gemma-4-31B-it-OptiQ-4bit", "contextWindow": 131072 },
+        { "id": "mlx-community/Qwen3-VL-32B-Instruct-8bit", "contextWindow": 131072 }
+      ]
+    },
+    "rapid-mlx": {
+      "baseUrl": "http://localhost:11437/v1",
+      "api": "openai-completions",
+      "apiKey": "none",
+      "models": [
+        { "id": "mlx-community/gpt-oss-120b-MXFP4-Q8", "contextWindow": 131072, "reasoning": true }
+      ]
     }
   }
 }

packages/local-models/README.md

@@ -12,14 +12,14 @@ Getting a **local model worth using** behind every harness in this repo.
 ./probe.sh                                   # hardware → tier
 # read MODELS.md § your tier, pick a model, then ONE of:
 ollama pull <model>                          # zero-friction
-serve/serve-llamacpp.sh model.gguf           # GGUF, full control, :8080
-serve/serve-vllm.sh <hf-id> <parser>         # production path, :8000
+serve/serve-llamacpp.sh model.gguf           # GGUF, full control, :11435
+serve/serve-vllm.sh <hf-id> <parser>         # production path, :11438
 ```
 
 ## The contract
 
 Everything serves **OpenAI-compatible** on fixed local ports — ollama
-`:11434`, llama.cpp `:8080`, vLLM `:8000` — and every harness preset in this
+`:11434`, llama.cpp `:11435`, vLLM `:11438` — and every harness preset in this
 repo already points at all three:
 
 | Harness | Where the endpoint is wired |

packages/local-models/serve/serve-llamacpp.sh

@@ -1,5 +1,5 @@
 #!/usr/bin/env bash
-# Serve a GGUF model via llama.cpp's OpenAI-compatible server on :8080 —
+# Serve a GGUF model via llama.cpp's OpenAI-compatible server on :11435 —
 # agent-ready flags (large context + native tool-call template).
 # Usage: serve-llamacpp.sh /path/to/model.gguf [ctx-size]
 set -euo pipefail
@@ -20,4 +20,4 @@ exec llama-server \
   --ctx-size "$CTX" \
   --jinja \
   --host 127.0.0.1 \
-  --port 8080
+  --port 11435

packages/local-models/serve/serve-mlx.sh

@@ -0,0 +1,20 @@
+#!/usr/bin/env bash
+# Serve an MLX model via mlx_lm.server's OpenAI-compatible API on :11436 —
+# Apple-silicon-native (fastest local backend on M-series). Deliberately :11436,
+# off llama.cpp's :11435, so both can co-tenant (matches the `mlx` harness provider).
+# Usage: serve-mlx.sh <hf-repo | mlx-serve-preset>   (default: gpt-oss-120b flagship)
+set -euo pipefail
+
+MODEL="${1:-mlx-community/gpt-oss-120b-MXFP4-Q8}"
+PORT="${PORT:-11436}"
+
+# Prefer the dotfiles launcher (presets: flagship/fast/code/vision/vlm/…).
+if command -v mlx-serve >/dev/null 2>&1; then
+  exec env PORT="$PORT" mlx-serve "$MODEL"
+fi
+
+command -v mlx_lm.server >/dev/null 2>&1 || {
+  echo "mlx_lm.server not found. Install: uv tool install --python 3.12 mlx-lm" >&2
+  exit 1
+}
+exec mlx_lm.server --model "$MODEL" --host 127.0.0.1 --port "$PORT" --trust-remote-code

packages/local-models/serve/serve-vllm.sh

@@ -1,5 +1,5 @@
 #!/usr/bin/env bash
-# Serve a HF model via vLLM's OpenAI-compatible server on :8000 — agent-ready
+# Serve a HF model via vLLM's OpenAI-compatible server on :11438 — agent-ready
 # flags (auto tool choice + parser). vLLM is the production-grade path
 # (throughput, concurrent harnesses); llama.cpp/ollama are the lighter paths.
 # Usage: serve-vllm.sh <hf-model-id> <tool-parser> [max-model-len]
@@ -20,4 +20,4 @@ exec vllm serve "$MODEL" \
   --enable-auto-tool-choice \
   --tool-call-parser "$PARSER" \
   --host 127.0.0.1 \
-  --port 8000
+  --port 11438