MLX Router — Local Inference Gateway

MLX Router — Local Inference Gateway

بِسْمِ ٱللَّهِ ٱلرَّحْمَـٰنِ ٱلرَّحِيمِ

Service: mlx-router (com.alai.mlx-router) Port: 11500 (127.0.0.1) Status: Production (2026-05-01) Owner: ALAI System Infrastructure MC: #10429

Overview

MLX Router is ALAI’s local inference gateway that routes AI inference requests to zero-cost MLX models running on ANVIL (Mac Studio M3 Ultra, 96GB). It provides tier-fallback routing: tier1 MLX (local, $0 cost) → tier2 FORGE Ollama ($0 cost) → tier3 Anthropic API (metered cost).

Purpose: Reduce inference costs by offloading read-only agent workloads to local MLX models. Anthropic API is reserved for tier3 fallback only.

Cost Savings: All MLX and FORGE requests logged at cost_usd=0. Estimated 95%+ reduction in inference costs for wired agents.

Architecture

flowchart LR
    subgraph Caller
        A[Agent/Client]
    end
    
    subgraph MLX-Router["mlx-router.js :11500"]
        R[Route by model_class]
    end
    
    subgraph Tier1["Tier 1: MLX Local (ANVIL)"]
        M1[classify → :11437<br/>Qwen3-8B-4bit]
        M2[code → :11438<br/>Qwen2.5-Coder-32B]
        M3[reason → :11435<br/>Gemma-4-26B]
        M4[audit → :11436<br/>Qwen3-32B]
    end
    
    subgraph Tier2["Tier 2: FORGE Ollama"]
        F[10.0.0.2:11434<br/>qwen3/deepseek/etc]
    end
    
    subgraph Tier3["Tier 3: Anthropic API"]
        C[claude-haiku/sonnet/opus]
    end
    
    A -->|POST /v1/chat| R
    R -->|Health: UP| M1
    R -->|Health: UP| M2
    R -->|Health: UP| M3
    R -->|Health: UP| M4
    R -->|Tier1 DOWN| F
    F -->|Tier2 FAIL| C
    
    M1 -.->|cost_usd=0| CT[cost-tracker.js]
    M2 -.->|cost_usd=0| CT
    M3 -.->|cost_usd=0| CT
    M4 -.->|cost_usd=0| CT
    F -.->|cost_usd=0| CT
    C -.->|metered| CT

Service Management

Start

launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.alai.mlx-router.plist

Stop

launchctl bootout gui/$(id -u)/com.alai.mlx-router

Restart

launchctl bootout gui/$(id -u)/com.alai.mlx-router
launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.alai.mlx-router.plist

Check Status

launchctl print gui/$(id -u)/com.alai.mlx-router
# Look for: state = running
# Get PID from output

View Logs

# Stdout (health probes, routing decisions)
tail -f /tmp/com.alai.mlx-router.stdout.log

# Stderr (errors)
tail -f /tmp/com.alai.mlx-router.stderr.log

Health Check

Endpoint

curl -s http://127.0.0.1:11500/health | jq

Expected Output

{
  "status": "ok",
  "endpoints": {
    "classify": { "available": true, "lastCheck": "2026-05-01T09:00:00.000Z", "latencyMs": 8 },
    "code":     { "available": true, "lastCheck": "2026-05-01T09:00:00.000Z", "latencyMs": 7 },
    "reason":   { "available": true, "lastCheck": "2026-05-01T09:00:00.000Z", "latencyMs": 3 },
    "audit":    { "available": true, "lastCheck": "2026-05-01T09:00:00.000Z", "latencyMs": 5 }
  }
}

Healthy state: All four endpoints show available: true, latency <50ms.

Unhealthy state: If available: false, that model_class will fall to tier2 FORGE on next request.

Model Classes

Note on latency: MLX inference is sequential and slow. 8B models take ~14s, 32B models take ~94-117s. Not suitable for synchronous user-facing work. Use for background/async agent tasks only.

Tier Fallback Chain

1. Tier 1 — MLX Local (ANVIL): 127.0.0.1 ports 11435-11438, cost=$0
   • Health-gated: If endpoint available: false, skip to tier2
   • Timeout: 120s
2. Tier 2 — FORGE Ollama: 10.0.0.2:11434, cost=$0
   • Models: qwen3:8b (classify), qwen3-coder:latest (code), deepseek-r1:70b (reason), qwen3:32b (audit)
   • Timeout: 60s
3. Tier 3 — Anthropic API: Metered cost
   • Models: claude-haiku-4-5 (classify), claude-sonnet-4-6 (code/reason), claude-opus-4-7 (audit)
   • Timeout: 60s

Fallback triggers: HTTP error, timeout, or health probe failure. Router tries tier1 → tier2 → tier3 until success or exhaustion.

Cost Verification

All MLX and FORGE requests log cost_usd=0 to cost-tracker.js.

# Check today's MLX costs (should be 0.0)
node ~/system/tools/cost-tracker.js summary today | grep mlx-local

# Query cost_events.db directly
sqlite3 ~/system/databases/costs.db \
  "SELECT backend, SUM(cost_usd) as total, COUNT(*) as requests 
   FROM cost_events 
   WHERE backend='mlx-local' 
   GROUP BY backend;"

# Expected: mlx-local | 0.0 | <count>

Adding a New Model Class

1. Add MLX endpoint to ~/system/tools/mlx-router.js in MLX_ENDPOINTS object:

   new_class: {
     url: 'http://127.0.0.1:11439',
     modelId: '/Users/makinja/system/research/mlx-models/NewModel-4bit',
     shortname: 'newmodel-mlx',
     maxConcurrent: 1,
   }

2. Add tier2 FORGE fallback in FORGE_FALLBACK:

   new_class: { model: 'forge-model:latest', url: 'http://10.0.0.2:11434' }

3. Add tier3 Anthropic fallback in ANTHROPIC_FALLBACK:

   new_class: 'claude-sonnet-4-6'

4. Extend capability table at ~/system/specs/mlx-capability-table.md with routing rationale.

5. Restart daemon:

   launchctl bootout gui/$(id -u)/com.alai.mlx-router
   launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.alai.mlx-router.plist

6. Verify health:

   curl -s http://127.0.0.1:11500/health | jq '.endpoints.new_class'
   # Should show available: true

Wiring an Agent

Add inference: block to agent’s YAML frontmatter in ~/system/agents/definitions/<agent>.md:

inference:
  prefer_inference: mlx-router
  model_class: classify
  router_url: http://127.0.0.1:11500/v1/chat
  rationale: "Read-only classification tasks — no production risk"
  wired_by: skillforge/MC#<id>/<date>

Sync to active agents:

~/bin/agent-definitions-sync.sh

Currently wired agents (2026-05-01): - sentinel-tester (classify) - sentinel-validator (reason) - sentinel-architect (audit)

Failure Modes

Performance Expectations

Latency (tier1 MLX): - 8B classify: ~14s (measured) - 32B code: ~117s (measured) - 32B reason: ~94s (measured) - 32B audit: ~120s (estimated)

Concurrency: - classify: 2 parallel requests - code/reason/audit: 1 request at a time (MLX is sequential)

Not for: - User-facing synchronous requests (too slow) - Real-time classification (<1s SLA)

Good for: - Background agent tasks (sentinel audit, QA checks) - Async workflows (overnight batch processing) - Read-only analysis (no Write/Edit risk)

Logs & Debugging

Daemon logs:

# Health probe output every 60s
tail -f /tmp/com.alai.mlx-router.stdout.log

# Example healthy output:
# [mlx-router] Health probe:
#   classify: UP (8ms)
#   code: UP (7ms)
#   reason: UP (3ms)
#   audit: UP (5ms)

Request routing logs:

# Each request logs tier used
# Example: [mlx-router] tier1 classify → qwen3-8b-mlx (470ms)
# Tier2/tier3 fallback logged with reason

Cost tracking:

# Every request creates a cost_events entry
sqlite3 ~/system/databases/costs.db \
  "SELECT model, backend, cost_usd, timestamp 
   FROM cost_events 
   WHERE backend='mlx-local' 
   ORDER BY timestamp DESC 
   LIMIT 10;"

Related Resources

• Capability Table: ~/system/specs/mlx-capability-table.md
• Ollama Fleet: ~/system/config/ollama-fleet.json
• LaunchAgent: ~/Library/LaunchAgents/com.alai.mlx-router.plist
• Cost Tracker: ~/system/tools/cost-tracker.js
• Agent Definitions: ~/system/agents/definitions/sentinel-*.md

MC History

• MC #10429: MLX router build + agent wiring (2026-05-01)
• MC #10391: SENTINEL v2 audit identified MLX orphans (2026-05-01)
• MC #10411: SENTINEL v3 decision item 2 = activate $0 inference

Last Updated: 2026-05-01 Status: Production Validation: Proveo 10/10 PASS