⚒ FORGE v1.0.0

your harness, your rules

Server

engine (27B)…

VRAM–

GPU util–

Model

Sampling

temperature top_p top_k min_p presence_penalty frequency_penalty repetition_penalty max_tokens seed (-1 = random) stop sequences (separate with ||)

Thinking

enable_thinking

preserve_thinking

reasoning_effort

Qwen warns: low effort in agent loops can cost MORE time overall via failed attempts and retries.

Agent

max turns bash timeout (s) tool output limit (chars)

auto-approve (YOLO — deny rules still apply)

workspace root enabled tools

System prompt

Replaces the entire system prompt. ~90 tokens by default — nothing hidden.

Permissions

bash rules (regex → action, last match wins)

bash default

saved ✓

Agent

Arena

Models

MCP

model fleet 0/0

⚒

FORGE — give the agent a task below

Enter send · Shift+Enter newline · Esc abort

settings apply per message — tune freely between turns

turn – speed – tok/s last gen – tok total out – tok context in – tok tool calls –

Fleet

Lanes run concurrently via vLLM continuous batching. A profile with an endpoint set can point at a second server (e.g. llama.cpp on :8001).

Select profiles, enter a prompt, hit Run.