Agent harnesses for local models — an overview

As of 11 October 2026. Which open harnesses turn a local model into an agent, how they differ, which to pick for what — and when it is better to assemble your own. A ranking of the top 20 out of 72 examined, with profiles, building blocks and an architecture for specific hardware.

⚠️ This overview goes stale within weeks. In 2026 some popular projects died or changed owners (Roo Code, Flowise, Continue, Aider), and new ones earned hundreds of thousands of stars in two months. I pulled the stars and releases from the GitHub API on 11 October 2026. I have not tested anything here on my own machine — this is a review of documentation, code, release notes and issues, not a measurement. Your measurement wins.

---

Table of contents

• Why this overview
• The shortest answer
• What a harness is
• Ranking: the top 20 harnesses
• Technical comparison
• Which one for which use case
• Verdict: build on a harness or assemble your own?
• Why they are not at the top
• How I computed the ranking
• Files in this directory
• Related course parts

---

Why this overview

I build multi-agent applications, local or hybrid:

• locally — up to 4 agents at once on DeepSeek-V4-Flash (4× AMD Radeon AI PRO R9700, llama-swap, 4 slots of 131k context each, about 13.6 tok/s per agent);
• hybrid — Claude Opus 5.5 as planner and orchestrator (the most intelligence), local agents as executors (the lowest cost of work).

Five use cases by which I rate every harness:

---

The shortest answer

• There is no single harness for everything. The best setup has three layers: a ready-made coding agent for U1, your own framework-based applications for U2 and U3, and a personal assistant with voice as the interface.
• U1 → OpenCode (or Goose): main agent on Opus 5.5, subagents on DeepSeek, permissions on command patterns. Do not build this yourself.
• U2 and U3 → your own application on Pydantic AI (the only framework with a built-in limit of 4 slots) or LangGraph (the most mature persistence) + building blocks from 02-BUILDING-BLOCKS.md. No code: n8n.
• U4 and U5 to start → Hermes Agent: voice notes on Telegram with local speech recognition (faster-whisper), a "vision bridge" for the text-only DeepSeek, the best cron. Note: speech synthesis goes through the cloud by default (Edge TTS), and a local Polish voice is not confirmed — a fully local voice with Polish Piper is offered by Moltis. Live conversation in Polish — Pipecat.
• DeepSeek-V4-Flash cannot see images and takes up practically all the VRAM — voice and vision must run on the CPU, on a freed-up card or in the cloud (03-ARCHITECTURE.md).
• Claude Code subagents on local models are not a supported feature. It works the other way round: Claude orchestrates, local agents are its tools (MCP).
• Local is not always cheaper. With short contexts Claude Haiku 5.5 (USD 0.10 / 0.50 per million tokens) costs about as much as this machine's electricity at full load; with contexts above 100k tokens and compared with Sonnet or Opus, local DeepSeek comes out clearly cheaper. Certain local advantages are privacy, no limits and working offline (costs).

---

What a harness is

A language model does nothing by itself — it answers with text. A harness is everything around it that turns it into an agent: the loop (think → call a tool → read the result → again), tools, context and memory management, permissions and sandbox, scheduling, channels (terminal, chat, voice), observability. You build the loop from scratch in course 1, part 6, and a long-task harness in course 2, part 3.

The ranking contains four kinds of harnesses:

---

Ranking: the top 20 harnesses

The columns "What it is", "What it is for" and "What problem it solves" answer the question of what each one beats the competition at. Details, local model configuration, weak points and ideas to steal — in 01-PROFILES.md.

---

Technical comparison

✅ present and works with a local model · ⚠️ partially or with a caveat · ❌ absent. "Hybrid" = whether one agent can run on Claude and others on your llama-swap. Stars and releases: GitHub API, 11 October 2026.

Before connecting any of them to llama-swap, read the seven pitfalls common to all — Responses API, DeepSeek's reasoning_content, the context window, prompt cache, number of tools, concurrency limit, tool schemas.

---

Which one for which use case

Rated 1–5 based on a documentation study (not a measurement). Bold — the best in the column. These numbers measure how well the features fit the use case and are not a recommendation: the "What to pick" table below also takes into account readiness for use (e.g. LangGraph has a 5 in U1, but for DevOps it is better to take the ready-made OpenCode than to build an agent on a framework).

What to pick — and with which building blocks:

How to put it all together on this hardware: 03-ARCHITECTURE.md.

---

Verdict: build on a harness or assemble your own?

Both — but in different layers.

What to learn from other people's code (readable core, good ideas): Pi — a minimal core and many embedding modes; nanobot — readable Python to fork; mini-swe-agent — an agent in 100 lines; OpenCode — a server with an API and agents as files; Hermes Agent — delegation with a result schema and cron watchers; ZeroClaw — the security model.

What to avoid: dead projects (Roo Code, Flowise, Continue, Plandex); Codex CLI with local models (Responses API only); LiteLLM without pinned versions (PyPI compromise on 24 March 2026); n8n and OpenClaw exposed to the internet.

---

Why they are not at the top

---

How I computed the ranking

1. Entry conditions (all four): public code and a license that allows self-hosting (source-available allowed, but marked ⚠️); a documented connection of a local model through an OpenAI-compatible endpoint (directly or — marked — indirectly, through an adapter); activity — a commit no older than 3 months and a release (including pre-release) no older than 6 months before 11 October 2026; the project is not officially discontinued — which is why Continue, for example, drops out: its maintainers announced the end of the project, although the last release dates from June 2026. Examined: 72 projects, 64 passed, 8 dropped out.

2. Six criteria, each 1–5, with weights according to my goals:

Score = (25·A + 20·B + 20·C + 15·D + 10·E + 10·F) / 5, that is 0–100. Ties are broken by C, then stars.

3. Scores of the top — change the weights and calculate for yourself:

The full list of 64 candidates with scores and the eight excluded — 01-PROFILES.md.

4. How I gathered the facts. Six research agents went through the documentation, READMEs, release notes and GitHub issues in parallel (separate ones for personal agents, coding agents, Python frameworks, TypeScript and workflow platforms, building blocks, and voice, vision and hybrid). I pulled the GitHub numbers myself from the API. The most surprising claims I checked at the primary source (among others the number of security advisories for OpenClaw and Open WebUI, the compromise of LiteLLM, the archiving of Flowise and Roo Code, the removal of Chat Completions from Codex CLI, the limiter in Pydantic AI, the attack on Mastra packages, the DeepSeek-V4-Flash model card, Anthropic pricing). Claims based only on blogs and marketing materials I mark in the profiles.

Limitations. Criteria A–F are an assessment, not a measurement. Stars are not quality (DeepSeek Harness collected 247k in two months). The U1–U5 scores come from documentation analysis. No harness was run on this machine.

---

Files in this directory

Related course parts

Measurements of this machine: local-llm-bench-results (speed and quality, also agents in parallel).

Text license: CC BY 4.0 — like the whole repository (LICENSE.md).