Skip to content
| Marketplace
Sign in
Visual Studio Code>Programming Languages>Nacho FlowNew to Visual Studio Code? Get it now.
Nacho Flow

Nacho Flow

Spicebox

|
7 installs
| (0) | Free
Active execution runtime for autonomous coding agents (Cline, Zoo Code, Cursor). In-flight loop defense, context compaction, and intelligent local/cloud model dispatch.
Installation
Launch VS Code Quick Open (Ctrl+P), paste the following command, and press enter.
Copied to clipboard
More Info

Nacho Flow

🌮 Nacho Flow: VS Code Extension

VS Code Marketplace Platform: VS Code | Cursor Bundled Pure Go Binary GitHub Stars License: MIT

The VS Code companion extension for Nacho Flow -- a hybrid model dispatcher and agent supervisor. Bundles the native Go binary directly. Zero CLI setup, zero Go toolchain required.

🌐 Website & Docs: spicebox.dev/nacho-flow · GitHub

Nacho Flow VS Code Extension - Live Analytics Dashboard, Sidebar Control Hub, and Cline Pairing


⚡ 60-Second Quickstart

  1. Open the Nacho Flow Sidebar: Click the 🌮 Nacho Flow icon in the VS Code Activity Bar.
  2. Launch the Engine: Under 1. Model Dispatcher, click Start. Status turns 🟢 Engine Online.
  3. Configure Your Agent: Under 3. Coding Agents, click Copy next to:
    • Base URL: http://127.0.0.1:8000/v1
    • Model ID: nacho-hybrid
    • (Optional API Key: sk-nacho-secret-key)
  4. Paste into Your Agent: Open Cline, Zoo Code, or Cursor settings, set Provider to OpenAI Compatible, paste the copied values.

Routine turns run on your local GPU at full speed, while complex reasoning automatically escalates to cloud frontier models.


✨ Features

🛡️ Agent Supervision

  • Cycle Killer: Kills repetitive N-gram loops in < 3s across reasoning, content, and tool lanes. Injects an immediate local override prompt to force a tool action. File writes are immune.
  • Kickstart (Read Slap & Thought Slap): Breaks analysis paralysis when agents get stuck in read-only loops (Read Slap) or hit max_tokens mid-monologue without emitting a tool call (Thought Slap). Injects a local steer prompt or escalates to a smarter tier (capped at 3 turns via anti-latching). Auto-suspends in Plan Mode.
  • Fairy Dust: Deploys frontier models (Claude Sonnet 5.5, DeepSeek R1) every N writes for quality checkpoints—without running heavy models continuously.
  • Tool Normalizer: Converts 8 malformed tool-call format families into standard OpenAI tool_calls JSON, eliminating 3-strike harness crashes from small local models.

🗜️ Nacho Token Saver (NTS)

  • Test Compaction: Strips passing test noise in-place across Go, Cargo, Pytest, and Jest at wire speed (0 allocs, < 10 µs). Preserves all failures, diffs, panics, and summaries.
  • Reasoning Pruning: Drops historical <think> blocks from older turns while keeping recent context fresh. Automatically migrates Anthropic cache breakpoints.
  • Pre-Routing Ingress: Compacts prompts before tier selection so inflated test outputs don't trigger premature cloud escalations.
  • Terminal Sanitizer: Resolves \r progress spinner spam in-place, purges ANSI escape codes, and deduplicates repeated tool/reminder loops.

🎛️ Sidebar Control Hub

  • Local / Remote Gateway: 1-click Start, Stop, Restart, and live streaming logs for the bundled Go engine. Connect across LAN or Tailscale with optional Bearer Auth.
  • Routing Profiles & Action Toolbar: Switch on the fly between built-in profiles (profile1.yaml, profile2.yaml, profile3.yaml) or create unlimited custom slots via + Add, ⧉ Clone, ✏️ Rename, - Delete, 🔍 Diff, and Reset with zero-downtime hot-reload:
    • Profile 1 (Standard Hybrid Pipeline — Default): Universal gateway configuration for autonomous coding agents (Cline, Kilo, Zoo, Cursor, Aider):
      • Core Coding Engine: Fast cloud workhorse (z-ai/glm-5.3-flash Low, 2–4s), Turn 0 Low Architecture Kickoff, Tier 2 Debug High, Tier 3 De-escalation Max ($0.65/M).
      • Kickstart Escalation Tier: Automatically resuscitates agents trapped in read-only browsing loops by injecting [SYSTEM OVERRIDE] and routing to google/gemini-3.8-flash to force concrete file edits or test executions.
      • Multimodal Vision Escapement Tier: Routes screenshot turns (Retries < 6 && HasImages) to google/gemini-3.8-flash for high-speed visual state analysis without bloating the primary code generator context.
      • Stream Defenses: Cycle Killer, Nacho Token Saver (NTS), and Thought Slap rescue integration.
    • Profile 2 (Deep Reasoning & Autonomous Benchmark): Tuned for deep chain-of-thought, complex algorithmic benchmarks, and runaway thought prevention:
      • Turn 0 Architecture Kickoff: Engages z-ai/glm-5.3-flash at Max Reasoning guarded by thought_slap: true. Rescues token budget truncations in 2.4 seconds into valid tool calls, completely preventing terminal stalls.
      • Tuned Cycle Killer: 16k thinking token ceiling, 12-word n-gram window, and 20 max repeats to tolerate dense ASCII tables and probability matrices without premature kills.
      • Fairy Dusting: Tactical Code Reviews at turn 15 (high reasoning) and Strategic Architecture Audits at turn 60 (max reasoning).
      • Proven Benchmark Record: 89 turns, 14.6 minutes, 93% test coverage, 0 race conditions, 100% pass rate on full Las Vegas Blackjack Monte Carlo simulation.
    • Profile 3 (Anthropic Native Benchmark): Direct first-party Anthropic Messages API pipeline (https://api.anthropic.com):
      • Zero intermediary proxy latency with native prompt caching (cache_control) and multimodal vision.
      • Claude Sonnet 5.5 Turn 0 Architecture → Claude Haiku 5.5 Fast Workhorse (Low & Medium reasoning, guarded by real-time test failure inspection !HasTestFail) → Claude Sonnet 5.5 Error Escalation → Claude Sonnet 5.5 High Reasoning → Claude Opus 5.5 On-Demand override.
      • Speed Record: Completed full autonomous project build in 39 turns in 8.3 minutes with a 100% test pass rate.
  • Provider Status: Real-time health chips for local engines (llama.cpp, Ollama, vLLM) and cloud APIs (OpenRouter, DeepSeek, Anthropic).
  • Agent Setup: 1-click copy for Base URL, API Key, and Model ID, plus 1-click Marketplace installers for Zoo Code, Cline, and Kilo Code.

🧠 Interactive YAML Config Assist

Edit any Nacho Flow profile or config.yaml with full IDE language intelligence:

  • Interactive CodeLenses: Click Ping Endpoint, N tier reference(s), Switch Provider, Choose Model, and Pick Rescue Tier directly above YAML blocks across providers, tiers, default_tier, kickstart.thought_slap, and fairy_dust.entries.
  • Live /v1/models Discovery & Autocomplete (Ctrl+Space): Queries live model lists from local (llama.cpp, Ollama, vLLM) and cloud (OpenRouter, Anthropic) endpoints with 60s caching, automatic ENV_* key resolution, and curated offline fallbacks.
  • Semantic Diagnostics, Quick Fixes & Pre-Flight Save Gate: Catches unknown provider/tier references, unreachable tiers after when: "true", invalid when: expressions, and tab indentation with 1-click Quick Fixes (Ctrl+.). Automatically pauses local engine restarts on save if validation errors exist.
  • LSP Navigation & Refactoring: Supports Go to Definition (F12), Find All References (Shift+F12), Rename Symbol (F2) across provider and tier references, and Document Outline (Ctrl+Shift+O).

📊 Real-Time Analytics Dashboard

Open with Ctrl+Shift+P -> Nacho Flow: Show Dashboard:

  • Financial Telemetry: Total spend, savings ($, %), local GPU turns, cloud turns, billed vs. avoided tokens. Filter by Today, This Week, All Time.
  • Live Route Inspector: Last 500 LLM requests -- token counts, latency, matched tier rule, provider, model, retry steps.
  • Auto-Tuner: Click Run Auto-Tuner to replay traffic.jsonl, find optimal tier thresholds, and apply a green/red YAML diff with 1-click selective apply.

🔥 Heat Seeker: Live Model Deals

Scans 300+ cloud models on OpenRouter for flash discounts and free endpoints:

  • Deal cards show discount %, prompt/completion pricing, tool support, and SWE-bench score.
  • 1-Click Adopt: Select a tier to replace, the extension backups config.yaml, updates the model, and hot-swaps with zero downtime.

🚦 Status Bar HUD

🌮 Nacho Flow ● Online | $193.74 Saved All Time (5% Local)

Hover for active profile, cost metrics, supervisor defense telemetry, and NTS stats. Click for a QuickPick menu: open dashboard, switch presets, start/stop/restart, open config.yaml.


🌶️ In-Chat Directives (@nacho:)

Override routing directly from your prompt in Zoo Code, Cline, or Cursor:

Directive Type Effect
@nacho:toggles Inspection Show live session switches
@nacho:status Inspection Show daemon uptime, spend, savings
@nacho:reset Management Hard reset session & restore defaults
@nacho:kickstart-off / on Switch Suspend / resume Kickstart
@nacho:cyclekiller-off / on Switch Suspend / resume Cycle Killer
@nacho:shield-off / on Switch Suspend / resume tool-call synthesis
@nacho:raw-on / off Switch Enable / disable raw SSE stream
@nacho:fairydust-off / on Switch Suspend / resume frontier checkpoints
@nacho:local Single Turn Force Local GPU tier
@nacho:cloud Single Turn Force Cloud Fallback tier
@nacho:reasoning Single Turn Force DeepSeek R1 / o3

⌨️ Command Palette

All features via Ctrl+Shift+P / Cmd+Shift+P:

Command Identifier
Show Dashboard nacho-flow.showDashboard
Open Controls in Sidebar nacho-flow.openSettings
Open Configuration (config.yaml) nacho-flow.openConfig
Switch Profile nacho-flow.switchProfile
Add New Profile (+) nacho-flow.addProfile
Clone Active Profile (⧉) nacho-flow.cloneProfile
Rename Active Profile (✏️) nacho-flow.renameProfile
Delete Profile (-) nacho-flow.deleteProfile
Compare Profile with Factory Template nacho-flow.compareProfileWithTemplate
Reset Profile to Factory Default nacho-flow.resetProfileToDefault
Switch Provider (CodeLens) nacho-flow.pickProvider
Choose / Pick Model (/v1/models) nacho-flow.pickModel
Pick Rescue Tier (CodeLens) nacho-flow.pickRescueTier
Ping Provider Endpoint (CodeLens) nacho-flow.pingProviderEndpoint
Find Provider References (CodeLens) nacho-flow.findProviderReferences
Run Auto-Tuner & Optimize Routing nacho-flow.runOptimizer
Refresh Heatseeker Deals nacho-flow.refreshDeals
Reset Circuit Breakers nacho-flow.resetCircuit
Refresh Statistics nacho-flow.refreshStats
Set Bearer Auth Token nacho-flow.setAuthToken

⚙️ Extension Settings

Setting Default Description
nachoFlow.engineMode "local" 'local' runs the embedded daemon; 'remote' connects to an external gateway
nachoFlow.daemonUrl http://127.0.0.1:8000 URL of the local Nacho Flow daemon
nachoFlow.remoteDaemonUrl http://192.168.0.205:8000 URL of the remote Nacho Flow daemon server (LAN / Tailscale)
nachoFlow.authToken "" Optional Bearer Auth Token for protected remote gateways
nachoFlow.autoStartDaemon true Auto-resume the local binary on VS Code launch if previously running
nachoFlow.localVramGB 16 Local GPU VRAM limit in GB for auto-tuning routing thresholds (0 for unconstrained)

📚 Documentation

  • Official Website: spicebox.dev/nacho-flow
  • Extension User Guide: spicebox.dev/nacho-flow/docs.html?doc=extension_guide
  • Core Engine & CLI Guide: spicebox.dev/nacho-flow/docs.html?doc=user_guide
  • GitHub Repository: github.com/dixieflatline76/nacho-flow

🌮 Support Nacho Flow

If Nacho Flow saved your sanity or your API bill:

  • Grab a copy of Spice on the Mac App Store or Microsoft Store -- you get a native desktop utility and it funds continued development.
  • On Linux? Star Nacho Flow on GitHub.

📄 License

  • The VS Code Companion Extension is open source under the MIT License.
  • The core Nacho Flow gateway daemon is licensed under GNU AGPL-3.0 with API Interoperability Exception.
  • Contact us
  • Jobs
  • Privacy
  • Manage cookies
  • Terms of use
  • Trademarks
  • Your Privacy Choices
  • Consumer Health Privacy
© 2026 Microsoft