DeepSeek Harness hit 100K stars in two days, turning "agent harness" from testing jargon into a category: the execution scaffolding wrapped around models is becoming a selection criterion alongside the models themselves. This piece puts the four most notable contenders on one table: DeepSeek Harness (dsh), Anthropic Claude Code, OpenCode, and OpenAI Codex CLI. Verdict first: this is not a one-of-four substitution - there are three distinct needs (out-of-box productivity, model-agnostic daily use, hackable foundation), each with its own best answer.
Boundaries: a representative comparison based on official repos and docs (as of 2026-08-15), not a benchmark under identical conditions; star counts are GitHub API snapshots; features move fast - official docs are authoritative. Background in the Harness hotspot; using DeepSeek models with Claude Code is covered in the DeepSeek × Claude Code SOP.
1. The Overall Table
| Dimension | DeepSeek Harness (dsh) | Claude Code | OpenCode | Codex CLI |
|---|---|---|---|---|
| Positioning | Composable agent runtime (foundation) | Terminal coding agent (product) | Terminal coding agent (open daily use) | Terminal coding agent (OpenAI official) |
| GitHub stars | 101,905 (two days old) | 141,500 | 197,583 | 106,006 |
| Language | TypeScript | Python (CLI distribution) | TypeScript | Rust |
| License | MIT | Repo open, service closed | Open source | Apache-2.0 (CLI) |
| Model binding | None (model is a plugin) | Claude family (swappable via compat layer) | None (multi-provider) | OpenAI family mostly |
| Extension model | Everything is a plugin (kernel to UI swappable) | Skills/MCP (edge extension) | Plugins/provider config | Limited config |
| Observability | Append-only trajectory: resume/fork/replay | Session logs | TUI session management | Session logs |
| Interface | Local Web UI (127.0.0.1:3080) | Native terminal | Terminal TUI | Native terminal |
| Maturity | v0.1 developer preview (breaking-change warning) | Commercial mature | Community mature | Commercial mature |
On the snapshots: OpenCode's 197K stars are the highest stock (long-accumulated community project); dsh's 100K in two days is a speed record of hype, not a stock comparison; Claude Code and Codex stars include brand effects - real capability per official docs.
2. Best Range, One by One
DeepSeek Harness: the foundation for people who modify agents. It wins on "decomposability": model, tools, sandbox, loop, orchestration, UI are all Cordis plugins - swap any layer without forking; Minimal mode is the environment for official model benchmarks, directly reusable for evaluation. It loses on maturity: v0.1 preview, breaking-change warnings, ~300 unreviewed plugins, spartan UI. Verdict: for builders, researchers, and teams wanting a self-hosted fully open stack (MIT weights + MIT runtime); not for people who just need to get work done today.
Claude Code: the out-of-box productivity ceiling. Native terminal experience, the most mature task orchestration and skills ecosystem, plus Claude-family long-task reliability - the representative of "pay for certainty." Costs: deep binding to Anthropic models and subscription; using cheaper DeepSeek models requires the Anthropic-compatible layer (we have an SOP). Verdict: teams and heavy users with budget who want maximum completion rates.
OpenCode: the model-agnostic open-source daily driver. 197K stars, TypeScript, multi-provider plug-and-play, the highest community stock of the four. It and dsh are both open source but on different roads: OpenCode is "a better coding agent product"; dsh is "a runtime any agent can grow out of." Verdict: if you refuse vendor lock-in and want a stable usable open terminal agent, this is it.
Codex CLI: the OpenAI-ecosystem default. Rust implementation, Apache-2.0, deep synergy with OpenAI models and cloud sandbox. If your workflow already lives inside the ChatGPT/Codex subscription ecosystem, it's the smoothest line; cross-vendor flexibility is the weakest of the four. Verdict: the default answer for OpenAI subscribers.
3. Layered Selection and Combinations
| Your need | Pick | Combination tip |
|---|---|---|
| Max productivity today, budget no object | Claude Code | Pair with skills to bank team assets |
| Open-source daily use, multi-model cost saving | OpenCode | DeepSeek-V4 as workhorse, Claude for hard problems |
| Building/modifying your own agent, or self-hosting fully open | DeepSeek Harness | MIT weights + dsh runtime, end-to-end no proprietary dependency |
| Already in the OpenAI ecosystem | Codex CLI | With Codex cloud sandbox |
A three-layer lens makes the rivalry clearer: the model layer is commoditizing (replaceable), and the moat migrates to "the infrastructure that runs the models"; dsh turns the orchestration paradigm into composable blocks, but the paradigm itself (loop, tools, trajectory) doesn't break new ground; for ordinary developers the realistic future is multiple harnesses coexisting - work in Claude Code, save money with OpenCode, watch dsh grow up.
FAQ
Q1: Which of the four is "strongest"? A1: No single answer. For out-of-box productivity and task completion, Claude Code is the most commercially mature; for open-source stock and model-agnostic daily use, OpenCode (197,583★) is the steadiest; for architectural composability and a self-hosted open stack, dsh is the boldest; for OpenAI ecosystem synergy, Codex is the smoothest. "Strongest" depends on whether you are a user, a saver, a builder, or an ecosystem subscriber.
Q2: Can DeepSeek Harness replace Claude Code? A2: Not in the short term. dsh is a v0.1 developer preview (official breaking-change warning), unreviewed plugins, spartan UI, deployment friction. Its rival is not Claude Code's present but the "composable agent runtime" category's future. To save money with DeepSeek models on a mature harness today, OpenCode or Claude Code + compatibility layer is steadier.
Q3: Why is dsh's "everything is a plugin" different from ordinary plugin systems? A3: Ordinary tools (including Claude Code/OpenCode/Codex) extend at the periphery - add tools, connect MCP, install skills - while the kernel loop, orchestration, and UI are welded shut. dsh is built on the Cordis meta-framework where models/tools/skills/sessions/sandbox/storage/loop/orchestration/UI are all plugins; replacing any layer requires no framework-source changes - that's the difference between "edge extensibility" and "whole-system composability."
Q4: OpenCode and DeepSeek Harness are both open source - how to choose? A4: Decide whether you want a product or a foundation. OpenCode is a polished terminal coding agent - install and use, multi-provider switching, "daily driver" positioning. dsh is a runtime foundation - its value is decomposability and researchability (Minimal mode for model evaluation, Creator mode for plugin experiments), "builder" positioning. Most developers should pick OpenCode today and re-evaluate dsh when its API stabilizes.
Q5: Does harness choice affect how to read model benchmark scores? A5: Yes, enormously. DeepSeek officially runs its Code Agent benchmarks in Harness Minimal mode (just bash + file editing) - the same model can differ by a tier in production depending on harness and tooling scaffolding. Whenever you see "XX model #1 at coding," ask first: measured in which harness, which mode?
References
- GitHub API snapshots (2026-08-15): deepseek-ai/deepseek-harness 101,905★/MIT/TypeScript; anthropics/claude-code 141,500★; anomalyco/opencode 197,583★/TypeScript; openai/codex 106,006★/Rust
- DeepSeek official: Harness launch page (everything is a plugin, four modes, Minimal & benchmarks)
- NetEase Tech / Eigent AI / AI Reading Hub (InfoQ): dsh launch coverage and analysis (2026-08-13/14)
- Official docs of each project: Claude Code, OpenCode, Codex CLI features and licenses
A representative comparison (2026-08-15), not a hands-on benchmark; stars and features per official sources. Related: Harness hotspot | Harness quickstart SOP | DeepSeek × Claude Code SOP | Frontier coding model comparison