DeepSeek Harness (dsh) is DeepSeek AI's open-source agent harness, still in developer preview, and HagiCode is a workbench that runs dsh as one of its executors. Use dsh alone to try its plugin-based harness or script it; add HagiCode for parallel sessions, proposals and managed commits around it.
DeepSeek Harness, command dsh, is an open-source agent harness from DeepSeek AI. It runs as a local Web UI, a headless one-shot command, an SDK server or an ACP server, and almost every part of it is a plugin. HagiCode is a desktop and web AI programming workbench that drives agent CLIs, dsh among them, as its executor. It does not replace dsh; HagiCode starts dsh in its ACP mode and adds sessions, proposals and commit handling around it.
dsh is a developer preview whose own safety notice says it is experimental, unaudited and subject to breaking changes. If you want to explore the harness itself, write plugins or script it, stay with dsh alone. If you want to run it next to other agents with proposals and parallel sessions, HagiCode is the layer for that.
Open-source agent harness from DeepSeek AI, built from plugins on the Cordis framework; developer preview
Workbench that runs agent CLIs, dsh included, as executors
Local Web UI (dsh web), headless, SDK over stdio JSON-RPC, ACP; needs Node.js 22.19+ or 24+
Desktop app (Windows, macOS, Linux), Docker Compose container, Microsoft Store edition
MIT-licensed and free to run; you pay your model provider
Desktop and Container free; Store edition free with optional one-time TurboEngine purchase
DeepSeek API key, catalog providers (Anthropic, OpenAI, Moonshot, Z.ai and more), custom OpenAI- or Anthropic-compatible endpoints
Hosts no models; sends provider and model to dsh; optional OmniRoute routing
Sessions in the Web UI; the agent can delegate work inside a session
Several sessions at once, each a named Agent with visible status
The agent maintains a plan inside a session; no proposal lifecycle
OpenSpec proposals: propose, validate, apply, archive (3 concurrent free, 32 with TurboEngine)
Runs shell commands in the workspace; its docs include a GitHub review guide
AI-composed commit messages with the repository locked during the commit; Git and GitHub integration
npx @deepseek-ai/dsh web, then add a model key under Settings, Models
Desktop install, plus dsh on PATH and a configured model
Developers who want an open, hackable harness and accept preview software
People running dsh next to other agents, with traceable specs
dsh treats the model adapter, agent loop and interface as plugins composed through Cordis. Each profile has its own plugin set and a cordis.patch.yml for overrides, and dsh plugin --profile <name> add <package> installs a plugin into a profile. If you want to change how an agent harness behaves, that is the point of the project.
dsh web serves the browser UI. dsh headless "run the tests" answers one task and exits, reads a task from stdin, can emit newline-delimited JSON events and can resume a session by id. The sdk profile serves JSON-RPC over stdio, with a Python SDK guide, and the acp profile serves editors and automation clients.
In Settings, Models you enter a DeepSeek API key, pick a catalog provider, or define a custom endpoint speaking OpenAI Chat Completions, OpenAI Responses or Anthropic Messages. Model changes apply on the next request, and keys are write-only after saving, stored in $DSH_HOME/.credentials.yaml.
The README warns of compatibility-breaking changes, SAFETY.md calls the software experimental and not production-ready, and releases are alpha tags. The repository was created in August 2026, so expect fast change; that candor is useful when you decide where to run it.
dsh headless or build on the SDK or ACP profiles;HagiCode runs dsh as a subprocess in its ACP profile, in the repository you open, and passes the provider and model you choose. HagiCode exposes no credential fields for dsh, so sign-in and model keys stay in dsh.
Nothing is locked in: dsh remains a normal install that you can run with dsh web or dsh headless at any time.
npm install -g @deepseek-ai/dsh, so a dsh command is on your PATH.dsh --profile acp and rejects a --profile in extra arguments; the startup timeout defaults to 45 seconds.Run several dsh sessions at once, each with its own context, and see which is running, waiting or done. dsh's own sub-agent delegation stays inside a session; HagiCode adds the layer above.
Start each task as a proposal, discuss the approach with dsh inside it, then let dsh implement within that scope. Because dsh is a preview, the written proposal also records what you changed when a later release behaves differently.
HagiCode writes the commit message from the diff and locks the repository during the commit, so dsh output ends up as a clean history. Code Server opens the files dsh touched in a browser editor.
Monospecs lists your sub-repositories in .hagicode/monospecs.yaml, so a cross-repo change reaches dsh as one proposal. Vault stores reference projects and notes you register once and reuse.
OmniRoute centralises model routing across agents. dsh tool calls show in the same typed view as every other agent's, and optional TokenMonitor puts today's tokens and cost in the status bar.
dsh is MIT-licensed and free to run. You pay your model provider, for example through a DeepSeek API key; see the provider's pricing page for rates.
The official README runs it with `npx @deepseek-ai/dsh web`, which starts a Web UI at http://127.0.0.1:3080, and needs Node.js 22.19+ or 24+. For HagiCode, install it globally with `npm install -g @deepseek-ai/dsh` so `dsh` is on your PATH.
No, by its own statement. The README says dsh is a developer preview with compatibility-breaking changes, and SAFETY.md says it is experimental, unaudited and not production-ready.
Yes. HagiCode starts dsh in its ACP profile, passes the provider and model you choose, and checks that `dsh --help` identifies DeepSeek Harness. Install dsh and configure a model first.
No. dsh is the executor; HagiCode is the workbench around it, and you can still run dsh directly.
Pick dsh alone to explore the harness, write plugins or script one-shot tasks. Add HagiCode when dsh is one of several agents and you want parallel sessions, proposals and managed commits.
Assistant IA de code localisé, avec confidentialité et efficacité renforcées
desktopHero.loading