8

8gent Code

The Infinite Gentleman — Autonomous Coding Agent

v0.6.0 Open Source (MIT) Local-First Zero Usage Caps

Press → or click to advance • ? for keyboard shortcuts

The Problem With AI Coding Tools

Every developer hits these walls eventually

Usage Caps

Claude Code, Cursor, Copilot all throttle heavy users. Hit a limit mid-flow and your momentum dies. $20/mo doesn't mean unlimited.

Cloud Dependency

Your code goes to someone else's servers. Air-gapped environments, planes, and privacy-sensitive work are blocked entirely.

Model Lock-in

Claude Code only uses Claude. Cursor picks for you. When a better model drops, you wait for the vendor to integrate it.

Static Intelligence

These tools never learn YOUR patterns. Same prompts, same mistakes, same suggestions. No self-improvement loop.

8gent = The Open Alternative

Same agent power. No limitations. Gets smarter over time.

Unlimited Local Use

Run on Ollama with zero API costs. 14B parameter models on Apple Silicon. No caps, no throttling, ever.

Code Never Leaves Your Machine

Full offline mode. Works on planes, air-gapped networks, classified environments. Your code is yours.

Any Model, Instantly

Ollama (local), OpenRouter (cloud), LM Studio (custom). New model drops? ollama pull and go.

Self-Improving Engine

Autoresearch benchmarks + RL kernel fine-tuning. 8gent literally gets smarter the more you use it.

v0.6.0 By the Numbers

Shipped 2026-03-17

33

Packages

5

Apps

39

Benchmarks

40+

Slash Commands

5-0

H2H vs Claude

97%

Token savings with AST-first code nav

$0

Cost for unlimited local inference

15

Professional domain battle tests

5

Fine-tuned Eight model variants

Architecture Overview

33 packages, 5 apps, infinite gentleman energy

Core Agent

packages/eight — Main brain
packages/ai — Vercel AI SDK
packages/providers — Model router
packages/tools — Web, PDF, image
packages/toolshed — Discovery

Intelligence

packages/ast-index — 97% token savings
packages/planning — BMAD method
packages/kernel — RL fine-tuning
packages/memory — Persistent context
packages/validation — Evidence

Surfaces

apps/tui — Rich terminal UI
apps/clui — Desktop overlay
apps/dashboard — Admin panel
packages/voice — STT/TTS
packages/auth — Clerk OAuth

Competitive Landscape

8gent owns the Open + Local + Unlimited quadrant

Feature	8gent	Claude Code	Cursor	Aider
Local-first	Yes	No	No	Yes
No usage caps	Yes	No	No	Yes
Model freedom	Any model	Claude only	Multi	Any
BMAD methodology	Yes	No	No	No
Self-improving (RL)	Autoresearch + Kernel	No	No	No
AST-first nav	97% savings	Partial	No	No
Rich TUI	Animations, ADHD, Kanban	Plain CLI	IDE	Plain CLI
Desktop overlay	Tauri (5MB)	Planned	IDE	No
Voice I/O	Whisper + TTS	No	No	No
Open source	MIT	Partial	No	Apache

SWE Benchmark Results

Head-to-head: Eight (local 14B) vs Claude stand-in — LLM-judged, no string matching

Task	Difficulty	8gent	Claude	Winner
Async Mutex	Hard	95	85	8gent
Memory Leak	Hard	95	85	8gent
LRU Cache	Expert	95	85	8gent
Task Queue	Expert	85	40	8gent
State Machine	Expert	95	80	8gent

5-0

Perfect sweep

93.0

8gent avg

75.0

Claude avg

Self-Improvement Pipeline

Karpathy-style autoresearch + MetaClaw RL fine-tuning

Autoresearch Loop

Run 39 benchmarks with current prompt
Compare scores to baselines
Identify weak benchmarks
Generate enhanced prompt patterns
Validate improvement (LLM judge)
Keep if better, discard if worse
Loop forever

MetaClaw Kernel

Layer 1: Base Model (qwen3:14b)
Layer 2: Eight LoRA (centralized)
Layer 3: Personal LoRA (your patterns)
—
GRPO training pairs from sessions
Gemini Flash judge scores responses
MadMax: trains during idle/sleep
Never blocks active coding

Five Interaction Surfaces

Use 8gent however you work best

TUI

Rich terminal UI with 9 ASCII animations, ADHD/bionic reading mode, kanban board, ghost suggestions, Ctrl+Space voice input.

npm i -g @podjamz/8gent-code && 8gent

CLUI Desktop

Tauri 2.0 native overlay. Alt+Space from any app. Multi-tab sessions. 5MB binary. Permission server for tool approval.

tauri dev

CLI

Scriptable command-line agent. Pipe in prompts, get structured output. Auth commands, version info, headless mode.

8gent "fix this bug"

Voice

Whisper STT (local, no cloud). Hold Ctrl+Space to speak. TTS completion announcements via macOS Ava voice.

/voice record

Dashboard

Next.js admin panel. User management, session monitoring, usage charts, system health. Clerk auth + RBAC.

apps/dashboard

Roadmap

5 phases — 3 shipped, 2 in progress

✓

Phase 1: CLUI Desktop Overlay

Tauri 2.0 native app, Alt+Space toggle, multi-tab, permission server. SHIPPED v0.6.0

✓

Phase 2: Auth + User Accounts

Clerk OAuth, device code flow, Keychain storage, Convex DB, anonymous mode. SHIPPED v0.6.0

●

Phase 3: Full Voice Integration

Multi-language STT, noise filtering, accent support, latency optimization. IN PROGRESS

✓

Phase 4: Continuous Fine-Tuning (Kernel)

MetaClaw RL, GRPO training, Gemini Flash judge, MadMax scheduling. SHIPPED v0.6.0

●

Phase 5: Control Plane + Multi-Tenant

Stripe billing, tenant management, subdomain routing, usage enforcement. IN PROGRESS

○

Phase 6: Enterprise + Platform

SAML SSO, RBAC, audit logs, skill marketplace, model registry, API for embedded 8gent. PLANNED

Revenue Model

Free wedge, Pro upgrade, Enterprise custom

Free

$0

Local-only inference
Full agent capabilities
No sync, no cloud
Community support

Pro

$19/mo

Cloud sync across devices
Usage analytics
LoRA model hosting
Priority OpenRouter routing

Team

$49/seat

Shared configs & permissions
Team dashboard
Admin controls + SSO
Priority support

Enterprise

Custom

On-prem / air-gap deploy
SAML, RBAC, audit logs
SLA + dedicated support
Custom integrations

∞

The Infinite Gentleman

Open source. Local first. Zero caps. Self-improving. The Linux to Claude Code's Windows.

github.com/8gi-foundation/8gent-code

8gent.app

Built by James Spalding • Powered by Eight