deepseek-harness vs penguin-harness
Two of the top projects in Agent and MCP starters, side by side: who leads on each score part and who each one suits, then setup, license, activity and what each review found.
deepseek-harness
Plugin-based agent harness from DeepSeek with a local Web UI
penguin-harness
Agent framework that evolves an agent's prompts, skills and tools against benchmarks
Verdict
deepseek-harness ranks higher, 70 to 61. Of the 5 score parts we measure for both, deepseek-harness leads on 3, penguin-harness on 1 and 1 is even.
- Adoptiondeepseek-harness leads by 51 points, 100 to 49: a wide gap.
- FreshnessEven at 100.
- Maintenancedeepseek-harness leads by 7 points, 100 to 93.
- Easy to runpenguin-harness leads by 33 points, 33 to 0: a wide gap.
- Agent-readydeepseek-harness leads by 85 points, 85 to 0: a wide gap.
Pick deepseek-harness if you are
Developers building agents through plugins, tolerant of breaking changes
- Starts with one npx command; no Docker or database mentioned
- Everything-is-a-plugin design on Cordis, with a dsh-plugin topic for discovery
Pick penguin-harness if you are
Engineers tuning agent prompts and skills, or self-hosting agents
- Five RSI algorithms run on the same benchmarks, scoreboard and per-version snapshots
- Desktop app, Web App, CLI and SDK share one engine and the ~/.penguin/data directory
| What we compare | deepseek-harness | penguin-harness |
|---|---|---|
| Score parts, out of 100 | ||
| Adoption | 100, widely used | 49, known |
| Freshness | 100, active | 100, active |
| Maintenance | 100, healthy | 93, healthy |
| Easy to run | 0, hard | 33, some setup |
| Agent-ready | 85, ready | 0, none |
| Facts from GitHub and the README | ||
| Stars | 257k | 2.9k |
| License | MIT (permissive) | Apache-2.0 (permissive) |
| Last commit | Oct 2026 | Oct 2026 |
| Last release | None published | Oct 2026 |
| Language | TypeScript | TypeScript |
| Docker | No | Yes |
| GPU | Not stated | Not needed |
| arm64 or Apple Silicon | Not stated | Mentioned |
deepseek-harness
DeepSeek Harness (`dsh`) is a TypeScript agent harness built on the Cordis plugin framework, where every component is a plugin. It runs via `npx @deepseek-ai/dsh web` or from a source checkout with pnpm, and serves a Web UI at 127.0.0.1:3080 by default. It is a developer preview, and the README warns of compatibility-breaking changes.
Who it is for: Developers building agents through plugins, tolerant of breaking changes
Strengths
- Starts with one npx command; no Docker or database mentioned
- Everything-is-a-plugin design on Cordis, with a dsh-plugin topic for discovery
- MIT license, documentation site and third-party license notices
- Web UI binds to 127.0.0.1 by default; --no-open skips the browser
Weaknesses
- Developer preview; README warns of compatibility-breaking changes
- No tagged release published
- README does not list supported models, providers or system requirements
- Safety notice must be read before running; its contents are not in the README
- TypeScript
- Needs Node.js, pnpm (source build)
penguin-harness
PenguinHarness runs agents with a small toolset and keeps the model fixed while it evolves the harness (prompts, Skills, tools, hooks) over scored rounds. Five RSI algorithms (default, OPRO, APE, ACE, AWM) ship as plugins and run on built-in Harbor benchmarks in Docker, with each candidate snapshotted as a version. It ships as a desktop app, Web App, CLI and TypeScript SDK, and can also generate agent apps from a prompt.
Who it is for: Engineers tuning agent prompts and skills, or self-hosting agents
Strengths
- Five RSI algorithms run on the same benchmarks, scoreboard and per-version snapshots
- Desktop app, Web App, CLI and SDK share one engine and the ~/.penguin/data directory
- Works with many providers plus custom OpenAI-protocol endpoints, including local models
- Agents can run on remote machines over ssh and drive Chrome through an extension
Weaknesses
- Benchmarks run as Harbor tasks in Docker, so self-evolution needs Docker
- Cost and accuracy claims come from the project's own suite, not independent runs
- RAM and disk requirements are not documented
- Company-level self-evolution and OpenShell integration are still unchecked roadmap items
- TypeScript, OpenAI, Anthropic, Google Gemini, OpenRouter
- Needs Node >= 24 (npm install), Docker (benchmarks), Model provider API key
- Docker
- env example file
More in Agent and MCP starters
Building with these?
We help teams pick an open-source AI stack and run it in production. Client work never changes a rank.