China AI Hub
Agents / DeepSeek Harness
DeepSeek Harness
Open-source agent harness from DeepSeek ('Everything is a Plugin') that powers its coding agent. All capabilities - models, tools, skills, sessions, sandbox, storage, loops, scheduling and UI - are composed from replaceable plugins. Developer preview; ships as CLI (dsh), Web UI, Electron Desktop app, Python SDK and ACP server.
DeepSeek Harness is a framework agent by deepseek, built on deepseek-v4-1-flash, deepseek-v4-pro, deployable self-hosted, and released open source.
| Key fact | Value |
|---|---|
| Company | deepseek |
| Type | framework |
| Framework | Cordis |
| Deployment | self_hosted |
| Open source | Yes (MIT) |
| GitHub | https://github.com/deepseek-ai/deepseek-harness |
| Documentation | https://deepseek-harness.github.io/deepseek-harness/en/guide/quickstart |
| Pricing | Software itself free and open source. Model usage billed by the configured provider (DeepSeek API pay-as-you-go: flash $0.15-$0.30/M input cache-miss, $0.60-$1.20/M output, off-peak = half of peak; v4-pro $0.66-$1.32/M input, $1.98-$3.96/M output). |
Capabilities
| Capability | Supported |
|---|---|
| Tool calling | Yes |
| Browser use | Yes |
| Computer use | Yes |
| MCP support | Yes |
| Memory | No |
| Planning | Yes |
| Multi-agent | Yes |
| API | Yes |
Use cases
- Interactive coding-agent chat via Web UI
- Headless one-shot tasks via dsh --profile headless with a task prompt
- Embedding agents in Python programs (Python SDK)
- Programmatic multi-session automation over ACP (Agent Client Protocol)
- GitHub PR review via webhook overlay
- Custom profiles/plugins/presets for different agent compositions
Limitations
- Developer preview - compatibility-breaking changes expected; not security-audited, not production-ready (SAFETY.md)
- Sandboxing/approvals do not guarantee isolation - run untrusted work in a disposable VM or container
- Web server binds loopback only; trusted-host list required for LAN access
- No VS Code/editor extension found in repo; ACP is automation-only
- Image input only via deepseek-flash; deepseek-v4-pro is text-only
- OAuth providers (e.g. Codex) not yet supported; API-key providers only
- No built-in memory; third-party memory MCP servers are interoperability examples only
- sdk-minimal profile pins danger-full-access permissions by default
DeepSeek Harness is DeepSeek’s open-source agent harness (“Everything is a Plugin”) that powers its coding agent. Models, tools, skills, sessions, sandbox, storage, loops, scheduling and the UI are all composed from replaceable plugins on the Cordis framework. It is a developer preview that ships as a CLI (dsh), Web UI, Electron Desktop app, Python SDK and ACP server.
The software is free and self-hosted; model usage is billed by the provider you configure. DeepSeek API pay-as-you-go ranges from $0.15–$0.30/M input (cache-miss) for flash up to $1.98–$3.96/M output for V4-Pro, with off-peak at half price.
See the DeepSeek profile and the DeepSeek-V4-Pro model page.
Underlying models
Sources
Data interpretation
| Field | Value | Evidence type |
|---|---|---|
| Company | deepseek | Official |
| Type | framework | Official |
| Underlying models | deepseek-v4-1-flash, deepseek-v4-pro | Official |
| Open source | Yes (MIT) | Official |
| Deployment | self_hosted | Official |
| Capabilities | Tool calling, Browser use, Computer use, MCP support, Planning, Multi-agent, API | Official |
Evidence types: Official = vendor documentation or official pages. See the sourcing policy.
What is DeepSeek Harness?
Open-source agent harness from DeepSeek ('Everything is a Plugin') that powers its coding agent. All capabilities - models, tools, skills, sessions, sandbox, storage, loops, scheduling and UI - are composed from replaceable plugins. Developer preview; ships as CLI (dsh), Web UI, Electron Desktop app, Python SDK and ACP server.
Which models does DeepSeek Harness use?
DeepSeek Harness uses deepseek-v4-1-flash, deepseek-v4-pro.
Is DeepSeek Harness open source?
Yes — licensed under MIT.
How much does DeepSeek Harness cost?
Software itself free and open source. Model usage billed by the configured provider (DeepSeek API pay-as-you-go: flash $0.15-$0.30/M input cache-miss, $0.60-$1.20/M output, off-peak = half of peak; v4-pro $0.66-$1.32/M input, $1.98-$3.96/M output).
Where does China AI Hub get its DeepSeek Harness data?
From 4 sources, last verified 2026-09-20.