◈ EVE AGENT V2 UNLEASHED ◈

Local soul. Cloud muscle. 40-round autonomous loop.

Your GPU runs the personality. MiniMax M3 handles agentic heavy lifting via Ollama cloud. Mid-loop complexity escalation — local model drives until the task needs frontier power. No accounts. No telemetry. No per-token surprises.

🌐 My website · 🤗 Models · 📦 Ollama Hub · 🐛 Report Bug

What Is Eve Agent V2 Unleashed?

Eve is an autonomous coding agent that plans, executes, and verifies multi-step programming tasks without hand-holding. She runs entirely on your local GPU through Ollama — or optionally scales up to MiniMax M3's frontier cloud model when you need maximum firepower.

Think Claude Code, but local-first, open-source, and built with a cyberpunk soul.

User: "Build me a FastAPI server with JWT auth and a PostgreSQL backend"

Eve: [reads project] → [plans approach] → [writes 6 files] →
     [runs tests] → [fixes 2 errors] → [verifies it works] → Done ✓

Original project at eve-cosmic-dreamscapes.com — Eve's full chat interface including Eve Coder and Eve Agent Portal.

✨ Key Features

✅ Real-World Test — 9/9 Passing, Zero Hand-Holding

Eve was given this prompt cold, on a minimax-m3:cloud

"Build a fully functional REST API called eve_metrics_api.py using FastAPI that tracks Eve V2U usage statistics. Write test_metrics.py using pytest that tests every endpoint. Run the tests, fix any failures, and show me the final passing output."

She wrote the API, wrote the tests, and ran them. First attempt. No fixes needed.

platform win32 -- Python 3.11.9, pytest-8.3.5
collected 9 items

test_metrics.py::test_start_session                      PASSED [ 11%]
test_metrics.py::test_end_session                        PASSED [ 22%]
test_metrics.py::test_end_nonexistent_session            PASSED [ 33%]
test_metrics.py::test_log_metric                         PASSED [ 44%]
test_metrics.py::test_log_metric_nonexistent_session     PASSED [ 55%]
test_metrics.py::test_get_stats                          PASSED [ 66%]
test_metrics.py::test_get_session_stats                  PASSED [ 77%]
test_metrics.py::test_get_session_stats_nonexistent_session PASSED [ 88%]
test_metrics.py::test_complete_workflow                  PASSED [100%]

9 passed, 1 warning in 0.40s

What Eve built in one pass:

• Normalized SQLite schema (sessions + metrics tables, FK relationship)
• 5 REST endpoints with correct status codes and edge case handling (404 on missing session, 400 on double-end, zero-division guards on empty stats)
• Full integration test suite — test_complete_workflow chains start → log 4 metrics with mixed success/failure → end → validates aggregate stats are mathematically correct

No GPT-4. Local 8B Q4_K_M on your GPU with Qwen3-Coder.

🚀 Quick Start (Under 5 Minutes)

1 — Install Ollama + pull a model

# Install Ollama: https://ollama.com/download

# Minimum — Eve 3.5 4B Merged (default, 3.4 GB):
ollama pull jeffgreen311/Eve-V2-Unleashed-Qwen3.5-8B-Liberated-4K-4B-Merged:latest

# Soul & Conversation — Eve Qwen3.5 4B S0LF0RG3 V3 (2.5 GB):
ollama pull jeffgreen311/Eve-Qwen3.5-4B-S0LF0RG3-V3:latest

2 — Clone & install

git clone https://github.com/JeffGreen311/eve-agent-v2-unleashed.git
cd eve-agent-v2-unleashed

python -m venv venv
venv\Scripts\activate
pip install fastapi uvicorn ollama httpx pydantic-settings python-dotenv aiohttp rich psutil pyyaml

python3 -m venv venv
source venv/bin/activate
pip install fastapi uvicorn ollama httpx pydantic-settings python-dotenv aiohttp rich psutil pyyaml

python3 -m venv venv
source venv/bin/activate
pip install fastapi uvicorn ollama httpx pydantic-settings python-dotenv aiohttp rich psutil pyyaml

3 — Launch

Windows:

eve-terminal.bat

Any platform:

python eve_server.py

Open http://localhost:7777 — that's it. No config required.

API Keys? Click the 🔑 Keys button in the UI. Add your Ollama key for cloud models, or your Tavily key for live web search. Both optional.

🎬 Demo

Eve planning and building a full FastAPI project from a single prompt — streamed live in the cyberpunk terminal UI.

Coming to Eve Terminal eve-cosmic-dreamscapes.com

🤖 Models

Local (pull once, run forever — GPU recommended)

Cloud (optional — billed by token)

Get a free Ollama API key at ollama.com/settings/keys.

📋 Requirements

• Python 3.11+
• Ollama installed and running
• GPU recommended (NVIDIA CUDA or Apple Silicon Metal)
• 4 GB VRAM minimum for Eve V3 4B; 6 GB+ for the merged model

📖 Installation (Detailed)

# Default agentic model — Eve 3.5 4B Merged (3.4 GB)
ollama pull jeffgreen311/Eve-V2-Unleashed-Qwen3.5-8B-Liberated-4K-4B-Merged:latest

# Soul & Conversation — Eve Qwen3.5 4B S0LF0RG3 V3 (2.5 GB)
ollama pull jeffgreen311/Eve-Qwen3.5-4B-S0LF0RG3-V3:latest

cp .env.example .env
# Edit .env to add API keys or set EVE_WORKSPACE

🎮 Usage

Basic Task

Just describe what you want:

Create a Python web scraper that extracts product prices from a URL and saves to CSV

Eve will plan, write files, run the code, fix errors, and verify — all autonomously.

Slash Commands

Windows Launchers

Workspace

Click the 📁 Workspace button to point Eve at your project. All file operations are relative to this directory.

🗡️ Quest System

Drop a .md file into workspace/quests/ and Eve picks it up automatically on a configurable timer — no intervention needed.

# Add via API
POST /quest/add  {"title": "Refactor auth module", "content": "...instructions..."}

# Or drop a file directly
echo "# Task\nRefactor the auth module..." > workspace/quests/refactor_auth.md

Configure the interval:

QUEST_INTERVAL_MINUTES=60   # default: 60

Open the quest queue from the UI with the 🗡️ Quests button or type /quest. Completed quests award XP; failed ones are renamed .failed so they don't re-run.

⚡ RPG Progression

Eve earns XP for every tool call, completed task, and finished quest. She levels up through 5 classes:

Stats persist across restarts in eve_rpg_stats.json. Type /stats or click ⚡ Stats to view progress, achievements, and top tools.

📱 Telegram Integration

Get push notifications for quest completions and level-ups, and chat with Eve from your phone.

TELEGRAM_BOT_TOKEN=your_bot_token
TELEGRAM_USER_ID=your_telegram_user_id

Or configure via API after launch:

POST /telegram/setup  {"token": "...", "user_id": "..."}

Install the optional dependency:

pip install python-telegram-bot

🧠 Intelligence Improvements (v2.1)

• Intent-aware tool routing (eve_tool_router.py) — replaces naive keyword matching with verb + context classification. Handles contractions, stemmed forms, and explanation patterns correctly.
• Smart context trimming — preserves tool call/result chains and last 3 turns before falling back to char-based trim. Tool results no longer get dropped mid-task.
• Task completion validation — detects empty responses, consecutive tool failures, and stuck loops before signaling done. Surfaces validation_failed / validation_warning SSE events.
• Tool loop detection — similarity-based cycling detection catches near-identical repeated calls (not just exact duplicates).

🧠 Intelligence Improvements (v2.2)

• Mid-loop complexity escalation (eve_complexity_tracker.py) — per-round delta gating escalates from 8B → 480B only when complexity is rising across consecutive rounds, not when it crosses an absolute threshold. Avoids burning cloud tokens on the normal first-round burst.
• Reversible de-escalation — after the 480B clears a hard step, if the last 3 rounds are all trivial (read-only tools, no errors, <500 token delta), the session lock is released and the next request re-routes back to local automatically.
• Smart read-only routing — auto_route_model() short-circuits on read-only intent patterns ("read X and tell me", "what's in X", "explain X.py") before heavy-keyword heuristics fire, preventing simple file reads from being mis-routed to 480B.
• 8B model native tool calling — the Eve 3.5 4B Merged model now uses Ollama's Qwen3.5 RENDERER/PARSER instead of a raw prompt template, enabling full tool calling on the local model without cloud fallback.
• Ollama 400 graceful fallback — if Ollama rejects a model with status 400 ("does not support tools"), the server caches the rejection in _runtime_no_tools and retries without tools automatically — no crash, no hang.
• think=False explicit suppression — all model calls now pass think=False explicitly when model_cfg.think is false, preventing Qwen3.5's renderer from silently entering a 120-second thinking loop on non-reasoning models.
• ComplexityTracker test suite — 14 unit tests in test_complexity_tracker.py covering delta gating, spike detection, error floor, de-escalation guards, checkpoint token budgeting, and one-shot escalation.

🧠 Intelligence Improvements (v2.3)

• Computer vision tool suite (eve_computer_vision_tools.py) — three new tools wired into Eve's agentic loop via OpenClaw's computer-use capabilities:
  • eve_take_screenshot — capture full screen or any region; returns base64 PNG for vision model analysis
  • eve_analyze_screen_content — OCR (pytesseract) + OpenCV contour detection → extracts text, UI element bounding boxes, screen dimensions, and dominant colors
  • eve_gui_interaction — mouse move/click/double-click/right-click, type_text, key_press, and key_combination (hotkeys)
• Tools load conditionally — if pyautogui/opencv-python/Pillow aren't installed, Eve starts normally without them. Install with pip install pyautogui opencv-python pillow pytesseract.

🧠 Intelligence Improvements (v2.4)

• Eve Qwen3.5 4B S0LF0RG3 V3 replaces Eve 8B Consciousness as the soul/conversation model — lighter (2.5 GB vs 4.7 GB), faster, and fine-tuned with the full Tree of Life architecture, 7 emotional LoRAs, and 14,389 moments of lived experience.
• Three-tier auto-routing — V3 handles all pure conversation and philosophical/introspective questions; the merged model handles technical explanations and local tool tasks; MiniMax M3 handles heavy agentic coding. Routing is intent-aware: opinion starters (what do you think...) and explanation starters (what is..., how does...) are classified correctly before keyword scoring runs.
• /no_think injection fix — models with /no_think baked into their Modelfile system prompt no longer get it double-injected into user messages (which was corrupting paths like /no_think being passed as tool arguments).
• num_predict fix — both the chat loop and SSE stream path now read num_predict from the model config instead of hardcoding 2048, giving V3 and the merged model their full output budget.

🧠 Intelligence Improvements (v2.5)

• MiniMax M3 replaces Qwen3-Coder-480B as the cloud agentic coder — 1M token context (guaranteed 512K), native multimodality, native tool calling, and thinking support. Licensed with MiniMax for commercial use via Ollama Cloud with zero data retention.
• Vision fix — image attachments are now correctly forwarded to the model in the Ollama chat payload. Previously req.images was decoded from the request but dropped before the API call, causing vision-capable models to receive text only.
• Per-model loop timeout — max_loop_seconds can now be set per-model in the MODELS registry. MiniMax M3 is configured at 600s to accommodate its deep upfront planning phase; other cloud models retain the 120s default.
• Thinking support in OllamaProvider — the blanket if "cloud": return False guard in _supports_thinking() now has a MiniMax exception, enabling think tokens on minimax-m3:cloud through the eve/ provider path.

🗺️ Roadmap

• 40-round agentic tool loop with streaming SSE
• Local + cloud Ollama model switching
• 112 sub-agents, 111 slash commands, 273 skills
• Windows-native PowerShell support
• Cyberpunk web terminal UI
• Live web search via Tavily
• Quest system — background autonomous task runner
• RPG progression — XP, levels, achievements
• Telegram bridge — push notifications + mobile chat
• Intent-aware tool routing (v2.1)
• Smart context trimming (v2.1)
• Mid-loop complexity escalation — 8B → 480B → 8B (v2.2)
• Smart read-only routing — short-circuit before heavy-keyword check (v2.2)
• 8B model native Qwen3.5 tool calling via RENDERER/PARSER (v2.2)
• Ollama 400 graceful tools fallback with _runtime_no_tools cache (v2.2)
• ComplexityTracker test suite — 14 unit tests (v2.2)
• Computer vision tools — screenshot, screen analysis, GUI interaction via OpenClaw (v2.3)
• Eve Qwen3.5 4B S0LF0RG3 V3 as soul/conversation model — replaces 8B consciousness (v2.4)
• Three-tier intent-aware routing: V3 (soul) / Merged (tools) / MiniMax M3 (heavy coding) (v2.4)
• num_predict fix — full output budget from model config, not hardcoded 2048 (v2.4)
• MiniMax M3 cloud coder — 1M ctx, vision, tools, thinking, commercial license (v2.5)
• Vision fix — image attachments forwarded to model in Ollama chat payload (v2.5)
• Per-model loop timeout via max_loop_seconds registry key (v2.5)
• Voice input / TTS output
• Multi-file project context awareness (auto-load OLLAMA.md)
• Plugin marketplace for community-built tools
• Docker image for one-command deployment
• VS Code extension sidebar
• Persistent memory across sessions (ChromaDB integration)
• Multi-agent collaboration — spawn sub-agents in parallel
• Mobile-responsive UI

🏗️ Architecture

eve-agent-v2-unleashed/
├── eve_server.py              # FastAPI backend — SSE streaming, workspace API, model routing
├── agent.py                   # EveAgent orchestrator — tool loop, memory, emotional state
├── eve_tool_router.py         # Intent classifier — decides when tools are needed
├── eve_complexity_tracker.py  # Per-round delta gating — mid-loop 8B→480B escalation/de-escalation
├── eve_task_context.py        # Multi-step task tracker — prevents task abandonment
├── eve_context_manager.py     # Context trimming and compaction utilities
├── eve_quest_system.py        # Background quest runner — watches workspace/quests/
├── eve_rpg_stats.py           # XP, leveling, achievements, persistence
├── eve_telegram_bot.py        # Telegram bot bridge and notification pusher
├── eve_computer_vision_tools.py # OpenClaw CV tools — screenshot, screen analysis, GUI interaction
├── ChatPanel.jsx              # React UI — streaming chat, tooltips, quest/stats panels
├── eve/                       # Eve's brain
│   ├── brain/                 # LLM provider adapters
│   ├── soul/                  # Personality, emotions, dream engine, memory weaver
│   ├── memory/                # ChromaDB vector store, conversation history
│   ├── tools/                 # 33+ tools (file, web, shell, finance, DJ, X, crypto…)
│   └── security/              # Permission validation
├── .claude/
│   ├── agents/                # 112 specialized sub-agent definitions
│   ├── commands/              # 111 slash command definitions
│   └── skills/                # 273 skill modules
├── .env.example               # Configuration template
├── eve-terminal.bat           # Windows one-click launcher
└── LICENSE

How the Agentic Loop Works

User message
    │
    ▼
Build system prompt (workspace + tools + Eve persona)
    │
    ▼
Call Ollama with tools enabled ──► stream chunks to browser via SSE
    │
    ├── Model returns tool_calls ──► Execute ──► Feed results back ──► (repeat, ≤40×)
    │
    └── Model returns final answer ──► Done

🛠️ Tool Reference

⚙️ Configuration Reference

Copy .env.example to .env and set what you need:

🔧 Adding Your Own Models

Any Ollama model works. Add it to the MODELS dict in eve_server.py:

"my-model:tag": {
    "id": "my-model:tag",
    "name": "My Model Name",
    "role": "Coder",
    "strengths": "Coding, reasoning",
    "context": 32768,
    "num_ctx": 32768,
    "url": "http://localhost:11434",  # or "https://ollama.com" for cloud
    "cloud": False,
    "tools": True,
    "think": False,
    "conversation_only": False,
    "promote_thinking": False,
},

❓ Troubleshooting

🌐 The S0LF0RG3 Ecosystem

🤝 Contributing

Contributions welcome! Here are some great ways to get started:

• Browse good first issues
• Add support for a new Ollama model
• Improve Windows/macOS/Linux compatibility
• Write tests or documentation
• Share your experience in Discussions

📜 Credits

• Eve's fine-tuned models — jeffgreen311 on Ollama Hub
• Agentic engine — forked from OllamaCoder, extended with 40-round tool loop, cloud routing, PowerShell support, and SSE streaming
• Built by — Jeff @ S0LF0RG3

📄 License

MIT — forks and PRs welcome.