DSH vs Hermes vs Claude Code vs Codex vs OpenCode
Five capable agent systems, but they optimize for different jobs. This comparison separates source-backed product facts from practical trade-offs so you can choose by workflow rather than hype.
On this page
This is not a benchmark or a universal ranking. Product surfaces change quickly. Factual rows are grounded in first-party repositories/docs; “strengths” and “trade-offs” are practical interpretations of those documented architectures.
Quick answer
Choose DSH
If your priority is a plugin-first harness you can reshape, with a native local Web UI and DeepSeek's extensible runtime.
Choose Hermes
If you want a persistent personal agent: memory, self-improving skills, cron, remote execution and Telegram/Discord/other messaging are core features.
Choose Claude Code
If you want a focused, polished Claude-centered coding workflow in terminal/IDE with strong plugin and MCP customization.
Choose Codex
If you want OpenAI's coding ecosystem across local CLI, IDE, desktop and cloud with skills, MCP and automation.
Choose OpenCode
If your priority is open-source model freedom, a strong TUI and support for many cloud/local providers.
Feature matrix
| Area | DeepSeek Harness | Hermes Agent | Claude Code | OpenAI Codex | OpenCode |
|---|---|---|---|---|---|
| Primary identity | Plugin-first agent harness | Long-running personal / autonomous agent | Claude-centric coding agent | OpenAI coding agent across local + cloud surfaces | Open-source multi-provider coding agent |
| Main interfaces | Local Web UI, CLI/headless | TUI, Desktop, 20+ messaging platforms | Terminal, IDE, GitHub | CLI, IDE, desktop app, Codex cloud | TUI, desktop beta, client/server ecosystem |
| Model strategy | DeepSeek plus provider catalog and custom providers | Provider-agnostic: Nous Portal, OpenRouter, OpenAI and custom endpoints | Centered on Claude models | Centered on OpenAI / ChatGPT models | 75+ providers plus local models |
| Extension model | Everything-is-a-plugin architecture; MCP, Skills, subagents, workflows | Built-in toolsets, Skills, MCP, memory providers, integrations | Plugins, MCP, hooks, subagents | Skills, plugins, hooks, MCP, subagents, SDK / app server | Plugins, custom tools, MCP, agents/subagents, LSP |
| Persistent memory | Sessions and extensible capabilities; no Hermes-style self-learning memory loop documented | Core feature: persistent memory, session recall and self-improving skills | Project instructions and Claude Code customization; not positioned as a long-running personal-memory agent | Project instructions, skills and product memory features; coding workflow remains the center | Project/config-driven; not positioned around persistent personal memory |
| Automation / remote | Headless workflows, schedules and SSH/local Web access | Built-in cron, remote backends and messaging delivery are central | Automation through hooks/plugins/GitHub workflows | Local/non-interactive workflows plus cloud, GitHub Action and remote product surfaces | Client/server architecture, ecosystem clients and configurable agents |
| License | MIT | MIT | All rights reserved; Anthropic Commercial Terms | Apache-2.0 | MIT |
| Maturity note | Developer preview; breaking changes are explicitly expected | Broad actively developed agent platform | Established commercial coding product | Established OpenAI coding product with open-source CLI | Fast-moving open-source project; desktop and newer plugin surfaces still evolve |
Strengths and trade-offs
DeepSeek Harness (DSH)
DeepSeek · You want to build or customize the harness itself, prefer a local Web UI, and care about plugin-level composability.
Strengths
- Very explicit plugin architecture
- MIT open source
- Native local Web UI
- MCP, Skills, subagents and workflows fit one extensible runtime
Trade-offs
- Still in developer preview
- Compatibility-breaking changes are expected
- Rapidly growing third-party plugin ecosystem means quality and security vary by plugin
Hermes Agent
Nous Research · You want an agent that persists across sessions, works from messaging apps, schedules tasks and actively builds reusable memory/skills.
Strengths
- Persistent memory and cross-session recall are first-class
- Built-in cron and remote messaging gateway
- Broad provider flexibility
- Strong browser, media, delegation and MCP tool surface
Trade-offs
- Much broader operational surface than a coding-only agent
- More state, integrations and credentials to manage
- Its persistent-learning behavior is useful but requires more attention to what the agent stores and reuses
Claude Code
Anthropic · You mainly want a polished coding agent and already prefer the Claude model ecosystem.
Strengths
- Focused coding workflow in terminal and IDE
- Strong plugins, MCP, hooks and subagent customization
- Tight integration with Claude models and Anthropic's product stack
- Straightforward project-oriented developer experience
Trade-offs
- Not open source under a permissive software license
- Model choice is Claude-centered rather than general provider-agnostic routing
- Less suited than Hermes to being a persistent cross-platform personal automation agent
OpenAI Codex
OpenAI · You want OpenAI's coding stack across CLI, IDE, desktop and cloud, with local execution plus managed product surfaces.
Strengths
- Open-source CLI under Apache-2.0
- Local CLI plus IDE, desktop and cloud options
- Skills, MCP, hooks, subagents and automation surfaces
- Strong sandbox / approval model and ChatGPT integration
Trade-offs
- OpenAI / ChatGPT ecosystem is the center of gravity
- The product spans several surfaces, so the mental model is broader than a single local harness
- Less provider-agnostic than Hermes or OpenCode
OpenCode
Anomaly · You want an open-source terminal-first coding agent with maximum choice of models and providers, including local models.
Strengths
- MIT open source
- 75+ providers and local-model support
- Strong TUI plus client/server architecture
- Built-in LSP, MCP, agents/subagents and plugin support
Trade-offs
- Some newer plugin APIs are still beta
- Desktop app is still labeled beta
- Does not make Hermes-style persistent personal memory and cross-platform automation its core product identity
Which one should you choose?
| Your priority | Tool to consider | Why |
|---|---|---|
| Build a deeply customized agent runtime | DSH | Its architecture explicitly treats nearly everything as a plugin. |
| Persistent memory + scheduled / messaging automation | Hermes | Memory, cron, remote backends and 20+ messaging platforms are first-class product concepts. |
| Claude-heavy software engineering | Claude Code | It is the most direct path into Anthropic's Claude coding workflow. |
| OpenAI / ChatGPT coding across local and cloud | Codex | CLI, IDE, desktop, cloud, skills and OpenAI account integration live in one product family. |
| Use many providers or local models | OpenCode | Provider breadth and local-model support are core differentiators. |
First-party sources
- DeepSeek Harness: official repository · providers.
- Hermes Agent: official repository · official docs · security model.
- Claude Code: official repository · official docs · repository license file.
- OpenAI Codex: official repository · configuration reference.
- OpenCode: official repository · providers · agents · MCP.