SuperQode
The harness engineering framework for coding agents, optimized for local and open models.
Build your own harness. Run it with any local, open, or frontier model. Measure and improve the layer around your coding agents instead of renting a locked, vendor-tuned loop.
What's new
Latest Models, Smaller Core and Python-Extensible Harnesses
Harness control plane
Own the layer around the model
local, BYOK, ACP
read, edit, shell, MCP
local and adapters
test, eval, evidence
HarnessSpec routes everything
The model is replaceable. The harness is yours.
Why SuperQode exists
Three kinds of coding agents
Three kinds of coding agents dominate the market today. Each gets something right, and every one of them keeps the layer that decides how your agent behaves out of your hands.
Labs selling a proprietary model inside a proprietary harness.
What is good
Results can be excellent, because the model and harness are tuned deeply together.
The problem
It is a new lock-in. You cannot inspect, version, or change the closed loop that edits your code.
Proprietary harnesses sold as "model independence".
What is good
You get to choose the model provider, which solves part of the problem.
The problem
Swapping models is not harness independence. The harness was built around one model family, so every switch stays fragile.
Open harnesses with the parts that matter behind enterprise gates.
What is good
A real step forward: more of the loop is visible, and many providers are supported.
The problem
Openness alone is not control. The harness still has to be engineered, and the pieces you need sit behind paid features and services.
How SuperQode solves it
SuperQode turns the full agent loop into a portable, versioned HarnessSpec in your repository. You control model routing, tools, memory, sandbox policy, verification, and workflows. Drop the subscription lock-in and the scattered CLAUDE.md and AGENT.md files, and adopt one harness.yaml you own.
Own the harness. Choose the model. Measure what works on your code. Optimize continuously.
Read the detailed blog postThe workbench
Own the harness from your terminal
Connect a model, build and run a harness, and inspect every session in an interactive TUI built for local and open models.
Build your own harness
Define tools, permissions, model policy, memory, and checks in one YAML HarnessSpec that lives in your repository.
Optimized for local and open models
Tune the harness around Ollama, MLX, LM Studio, vLLM, SGLang, TGI, DS4, llama.cpp, and open model families.
Measure and optimize
Prove changes before you adopt them, with doctor, test, evals, evidence, SkillOpt, and model packs.
Concept
The harness is the product surface around the model
A model alone is not a coding agent. The harness decides what context it sees, which tools it can use, how it asks for approval, how memory works, how checks run, and how failures are measured.
Why use someone else's coding harness?
Proprietary coding tools tune the harness for their own models. SuperQode lets you tune it for your repository, your models, your memory choices, and your workflow constraints.
Build
define the harness
Measure
run checks and evals
Extend
add memory and tools
Optimize
improve the loop
Model route
local, BYOK, ACP
Tool policy
read, edit, shell, MCP
Memory
local and external adapters
Measurement
doctor, eval, evidence
Capabilities
Everything needed to own the coding harness
Practical surfaces for building, running, and improving a harness you can actually inspect.
Interactive coding TUI
Switch models, pick harnesses, and inspect sessions without leaving the terminal, with a compact status line and keyboard-first pickers.
HarnessSpec
A portable YAML specification for tools, permissions, imports, model policy, memory, and workflows.
Harness independence
The agent loop lives in your repo: inspectable, versioned, and yours to improve. No locked product harness.
Model choice
Use local models, BYOK cloud models, subscription runtimes for Codex, Grok, and Google Antigravity, or ACP coding agents through the same harness layer.
Harness Protocol v1
A versioned lifecycle and durable event envelope with conformance checks, portable session export, and Core, Python, and ACP reference adapters.
Python extensions
Extend the native Core with opt-in tools, TUI commands, skills, hooks, and providers through the Python Extension API and entry-point packages.
Local dynamic workflows with RLM
Run recursive local workflows over large logs, traces, and diffs. Evidence stays outside the prompt with replayable lineage on disk.
Airplane Mode
Prepare local search roots and no-network workflows when you want a strict local-first coding setup.
Memory and context
Keep memory local by default, connect adapters when needed, and make repo context searchable.
Optimization
Evaluate harness behavior, optimize skills, and tune model policy packs before changing daily workflows.
Integration
Expose harnesses through MCP, connect ACP clients, and run the daemon for tool and channel workflows.
Control
Make approvals, sandbox behavior, trust checks, and plugin policy explicit instead of hidden in the tool.
Preset catalog
Harness presets for every provider
Every install ships a full preset catalog: general workflows maintained by SuperQode, plus model-family harnesses tuned for each provider. Open the picker with :harness, or jump straight in with sq --harness <name>.
SuperQode
Model-agnostic harnesses that run with any connected model.
corerecommendedany modelMinimal native core with the four-tool default: read, write, edit, bash.
workbenchrecommendedany modelFull-featured workbench with the complete tool surface.
no-toolrecommendedany modelPure reasoning and chat with no repository tools.
benchmark-codingspecializedany modelAutonomous headless runs for benchmarks like Harbor and Terminal-Bench.
Moonshot AI
First-party global API route with automatic context caching.
kimi-codingrecommendedkimi-k3 → k2.7-codeMaintained family route tracking the validated stable Kimi release: 1M context, max reasoning, parallel tools.
kimi-k3-codingpinnedkimi-k3Frozen K3 preset retained for reproducible configurations.
Z.AI
First-party Z.AI general API with GLM-family policy tuning.
glm52-codingmodel presetglm-5.2 → 5.1 → 5Long-horizon GLM-5.2 harness: 1M context, max reasoning, parallel native tools.
glm-codingmodel presetglm-4.6Strong agentic coder with native tools for local or endpoint GLM.
Alibaba
Local coding on the Qwen-Coder family via Ollama.
qwen-codingmodel presetqwen3-coder (Ollama)Low-temperature local coding with native tool calls.
DeepSeek
Local DS4 stack with health checks and managed-server guidance.
ds4-codingmodel presetds4-localLocal coding tuned for compact JSON tool calls.
ds4-fast-localmodel presetds4-localLatency-focused variant for fast local iteration.
Local Gemma on Apple silicon through MLX.
gemma4-codingmodel presetgemma4 (MLX)Local coding tuned for strict JSON tool calls.
gemma4-no-toolmodel presetgemma4 (MLX)Local reasoning with no repository tools.
MiniMax
Long-context reasoning with measured tool behavior.
minimax-codingmodel presetminimax-m1Long-context coding and reasoning with medium reasoning effort.
:harness # picker: maintained families + your harnesses :harness all # complete keyboard-navigable catalog :harness customize kimi-coding # editable project-owned copy sq harness list --recommended # CLI catalog with visibility tiers sq --harness kimi-coding # direct model-aware activation
Meta-harness
One layer over every coding harness
Use SuperQode the way teams use Omnigent from the Databricks AI team: as a meta-harness. One common layer over Claude Code, Codex, Grok Build, Google Antigravity, OpenCode, and the harnesses you write yourself, all behind the same HarnessSpec contract.
Claude Code and Claude Agent SDK
Drive the local Claude Code CLI, or the Claude Agent SDK with your Anthropic API key.
:connect claudeCodex
Run Codex on your subscription sign-in, or through the codex-sdk runtime.
:codexGrok Build
Connect xAI's coding agent over ACP on an eligible Grok subscription.
:grokGoogle Antigravity
Use the signed-in agy CLI, or the SDK runtime with a Gemini API key.
:connect antigravityOpenCode and ACP agents
Attach any Agent Client Protocol agent as the executing harness.
:connect acp opencodeHarnesses you write yourself
Ship a complete harness as a Python package, or import an Omnigent agent spec.
sq harness import-omnigentsuperqode harness import-omnigent agent.yaml --output harness.yaml superqode harness validate --spec harness.yaml superqode harness run --spec harness.yaml --prompt "summarize this repository"
Existing Omnigent agent.yaml specs convert into a portable harness.yaml you own. Read the Omnigent compatibility docs
Workflows
One command for every workflow
Every step, from your first local run to optimizing and integrating a harness, is a single copy-ready command.
uv tool install superqode # or run without installing: uvx superqode
superqode local init --repo .
sq --harness superqode.local.yaml # sq is the new shortcut for superqodeLocal first
Tune the harness around the model you actually run.
Local model stack
Doctor, smoke, bench, warm, serve, stop, and hardware fit checks.
Open model policy
Model packs control temperature, context, tool behavior, and session limits.
Offline work
Airplane Mode prepares local search roots and strict no-network sessions.
Search and memory
Local code search, optional semantic search, and memory provider commands.
A coding harness framework optimized for local and open models.
Build your own harness, use any model, measure it, extend it, and optimize it.
FAQ
Frequently asked questions
What SuperQode is, how the harness layer works, and how to get started.