v0.1.1 · Latest Linux & macOS MIT

Your AI tools,
amplified.

NEXUS sits beneath your AI CLIs — routing micro-tasks to local hardware, syncing personas across every tool, and tracking spend across your entire stack. Define once, run everywhere.

120+ tokens/sec local

12+ built-in personas

4 AI CLIs supported

0 Go toolchain needed

One command to install

$ curl -sSL https://raw.githubusercontent.com/canoo/agent-nexus/main/install.sh | bash

Downloads pre-built binary + clones repo to ~/.config/nexus

Get started → View on GitHub · 5 ★

zsh — nexus

~ $ nexus

⚡ NEXUS Framework Manager v0.1.1

──────────────────────────────────

▸ Install NEXUS

Configure

Health Check

Uninstall NEXUS

j/k: navigate • enter: select • q: quit

──────────────────────────────────

NEXUS ▸ task: generate commit message

NEXUS ▸ routing to local compute plane

NEXUS ▸ delegate → qwen2.5-coder:1.5b [supervisor band]

NEXUS ▸ ✓ complete [4.2s] [0 cloud tokens used]

~ $

🔌

Bring your own tools

Claude Code, Gemini CLI, Kiro, Cursor, Codex — or whatever ships next month. NEXUS works with what you already use.

🖥️

Bring your own models

Cloud APIs, local Ollama, or both. Route tasks to the right tier based on complexity, not vendor lock-in.

🔗

One config, every tool

Personas, routing rules, and orchestration logic defined once — shared across all AI tools via symlinks.

📊

Observe everything

Track usage, costs, and routing decisions across your entire AI toolkit — not just one tool's silo.

What NEXUS does

One framework. Every AI tool.

⚡ Expert Orchestrator

The right specialist for every task.

NEXUS acts as the brain between your AI CLI and every specialist agent. Before doing any work, it scans your persona registry for the right expert — and asks if none exists.

NEXUS ▸ task: design auth schema
NEXUS ▸ scanning persona registry...
NEXUS ▸ delegate → engineering-backend-architect
NEXUS ▸ DONE. context preserved.

🧠 Persona Registry

12+ built-in specialists.

From backend architect to TUI developer to mobile builder. Each persona has precise trigger words, allowed tools, and domain expertise. Define your own in minutes.

backend-architect bubbletea-tui code-reviewer frontend-dev git-workflow mobile-builder + more

🖥️ Local Compute Plane

Route micro-tasks to your GPU.

The nexus-ollama MCP server routes structured tasks — commit messages, lint fixes, test scaffolds — to local Ollama models. Deep work stays in the cloud. Boilerplate doesn't.

Supervisor (1.5B)

120+ t/s

commit, boilerplate, tests

Logic (3B)

~75 t/s

lint-fix, refactor

🔗 One Config, Every Tool

Define once. Sync everywhere.

A single NEXUS install symlinks your personas and routing rules into Claude Code, Gemini CLI, Kiro, and any other CLI that ships. Update the persona file once — all tools see it instantly.

Claude Code Gemini CLI Kiro CLI ...and more

🔌 MCP Tools

Automatic local delegation.

Six tools available via Model Context Protocol. Your AI CLI calls them transparently — no context switching, no copy-paste.

ollama_commit_msg ollama_lint_fix ollama_logic_refactor ollama_boilerplate ollama_test_scaffold ollama_health

📊 Observability Dashboard v0.2.0

See what your AI stack costs.

Session logging, model routing decisions, cloud vs local spend — all surfaced in the TUI. Tokscale integration brings in data from 20+ AI CLIs. Coming in v0.2.0.

Session #47  qwen2.5-coder:1.5b  4.2s   $0.00
Session #46  claude-3-7-sonnet   12.1s  $0.031
Session #45  llama3.2:3b         3.1s   $0.00
──────────────────────────────────────────────
Month:  Cloud $2.41  ·  Local $0.00  ·  Saved ~$8.20

Persona Registry

Built-in specialists. Ready to deploy.

Each persona is a .md file that loads into any AI CLI via NEXUS. Swap them out, compose your own, or pull community personas from the marketplace (v0.4.0).

⚡

Expert Orchestrator

agents-orchestrator.md

The brain. Routes every task to the right specialist before picking up a tool.

trigger: any task

🏗️

Backend Architect

engineering-backend-architect.md

Scalable system design, database architecture, API development, cloud infra.

trigger: architect, design schema, API

🎨

Frontend Developer

engineering-frontend-developer.md

Astro, React, Vue, Svelte, Tailwind. Components, pages, layouts, DX.

trigger: build UI, component, page

📲

Mobile App Builder

engineering-mobile-app-builder.md

Native iOS/Android and cross-platform frameworks. App store to device.

trigger: mobile, iOS, Android, app

⬛

TUI Specialist

bubbletea-tui.md

Go terminal UIs with Bubbletea v2, Bubbles v2, and Lipgloss. Multi-screen flows.

trigger: TUI, terminal UI, bubbletea

🔍

Code Reviewer

engineering-code-reviewer.md

PR review, quality gates, security patterns, and constructive critique.

trigger: review, audit, check PR

🚀

DevOps / CI Agent

engineering-devops-ci-agent.md

CI/CD pipelines, containerisation, infrastructure automation, deployment.

trigger: CI, deploy, pipeline, Docker

🌿

Git Workflow Master

engineering-git-workflow-master.md

Atomic commits, branching strategy, clean history, PR hygiene.

trigger: commit, branch, rebase, merge

🔥

Firebase Agent

engineering-firebase-agent.md

Firestore, Auth, Functions, Hosting, and the full Firebase / GCP stack.

trigger: Firebase, Firestore, Auth

🗺️

Dev Roadmap

engineering-dev-roadmap.md

Project planning, issue triage, milestone scoping, and roadmap execution.

trigger: roadmap, planning, milestone

🔄

Toolkiit Migrator

toolkiit-migrator.md

Bootstraps and retrofits repos to Codelogiic toolkiit standards.

trigger: Initialize, Retrofit

🧪

UI Tester

ui-tester-agent.md

Playwright automation, screenshot capture, visual QA, regression testing.

trigger: test UI, screenshot, Playwright

Learn how personas work → · Browse on GitHub ↗

Local LLM Performance

Real numbers. Real hardware.

Benchmarks measured via JSON-RPC against the MCP server on an RTX 3050 Mobile (4GB VRAM). Every task type, every model, actual throughput.

Speed Benchmarks

RTX 3050 / 4GB VRAM

Model	Band	t/s
qwen2.5:0.5b <0.5 GB	micro	228–261
qwen2.5-coder:1.5b ~1.2 GB	supervisor	~122
gemma2:2b ~1.5 GB	supervisor	~84
llama3.2:3b ~2.0 GB	logic	~73
qwen2.5:3b ~2.0 GB	logic	~72
qwen:4b ~3.0 GB	logic	~28
qwen2.5:7b ~4.7 GB	heavy	~10

supervisor band logic band heavy band (12GB+ VRAM)

Hardware Presets

Full guide →

Hardware	VRAM	Speed
RTX 3050 Mobile	4 GB	70–120 t/s
M3 Base	8 GB UM	45–50 t/s
RTX 3060 / 4060	8 GB	40–60 t/s
M3 Pro 18GB	18 GB UM	25–30 t/s
RTX 4060 Ti 16GB	16 GB	35–55 t/s
RTX 4090	24 GB	30–70 t/s
RTX 5090	32 GB	61–100 t/s
M3 Max 48GB	48 GB UM	18–35 t/s
M3 Max 96GB	96 GB UM	8–20 t/s

Supervisor model: commit-msg, boilerplate, test scaffolds · Logic model: lint-fix, code refactor

MCP Test Results · RTX 3050 · Default Config

✓

ollama_commit_msg

qwen2.5-coder:1.5b · ~4.5s

Clean conventional commit

✓

ollama_boilerplate

qwen2.5-coder:1.5b · ~7s

Reasonable Express+Zod scaffold

✓

ollama_test_scaffold

qwen2.5-coder:1.5b · ~3s

Correct describe/it blocks

⚠

ollama_lint_fix

llama3.2:3b · ~6.5s

Fixed most errors; edge case garble

✓

ollama_logic_refactor

llama3.2:3b · ~3s

Nested loops → filter+map

✓

ollama_health

— · instant

Lists all available models

Roadmap

Where NEXUS is headed.

Every milestone is tracked on GitHub. Pick an issue and open a PR — contributions welcome.

v0.1.x — Foundation

Released

· Interactive TUI (Go/Bubbletea v2) — install, configure, health check, uninstall
· Persona registry (12+ built-in specialists)
· nexus-ollama MCP server with 6 delegation tools
· Symlink architecture — one config for every AI CLI
· Linux and macOS support

v0.2.0 — Observability

In Progress

· Session logging — which model, how long, estimated cost (#9)
· TUI dashboard — live view of routing decisions and model usage (#10)
· Cost tracker — cloud vs local spend; "what this would have cost" estimation (#11)

v0.2.1 — Tokscale Integration

In Progress

· Adapter that calls tokscale --json and parses into NEXUS structs (#22)
· Unified dashboard — merge Tokscale CLI data with NEXUS metrics (#23)
· Graceful degradation when Tokscale is not installed (#25)
· Coverage for 20+ AI CLIs: Claude Code, Gemini, Cursor, Codex, Copilot, and more

v0.3.0 — Dynamic Routing

Planned

· Auto-select model band based on task complexity (#12)
· Latency-based fallback — transparent failover between local and cloud (#13)
· Chain-of-models — multi-step pipelines: draft → review → apply (#14)

v0.4.0 — Ecosystem

Planned

· Cross-tool persona sync — NEXUS format → Gemini YAML, Claude MD, Kiro steering (#15)
· Persona composition — combine traits from multiple personas (#16)
· Plugin system — extensible MCP tools via YAML/JSON specs (#17)

v1.0.0 — Stable Release

Future

· Windows support, Docker, Homebrew (#18)
· Stable API — frozen interfaces, semantic versioning, migration guide (#20)

View milestones on GitHub → · Good first issues ↗

Live from GitHub

Recent Activity

View all commits ↗

Commits

Discussions

Get started now

Ship with the right expert
every time.

One install. Every AI CLI you use, improved. No Go toolchain needed.


curl -sSL https://raw.githubusercontent.com/canoo/agent-nexus/main/install.sh | bash

Read the docs → Star on GitHub ★

Your AI tools, amplified.