DeepSeek Harness Dev Stack
AdvancedDeepSeek's open-source, plugin-built agent in a local web interface or desktop app, running DeepSeek, Kimi, GLM, Claude, or GPT.
Published 6 October 2026
About DeepSeek Harness Dev Stack
DeepSeek Harness is DeepSeek's open-source, MIT-licensed coding agent, and nearly every part of it is a swappable plugin: the model adapter, the tools, the session log, the agent loop, the sandboxes, and the interface are composed on the Cordis framework, so a developer who wants a different loop or sandbox replaces that piece instead of forking the agent. Out of the box it reads and edits the workspace, runs commands inside an operating-system sandbox, keeps a plan, hands subtasks to subagents, and asks for approval under a permission policy. It ships as a developer preview, and the project warns of compatibility-breaking changes between releases.
DeepSeek is the default and the first-class model route: one API key, a direct adapter with reasoning controls, and some of the lowest per-token prices among capable coding models, doubled during weekday peak hours. Kimi, GLM, Claude, and GPT come from the bundled provider catalog with their own API keys, and a relay, company gateway, or model server of your own joins as a custom endpoint, so each session can run a different model.
Most developers run the harness on their own computer, which has to stay on while it works. Its web command serves the interface to a local browser, desktop apps for Apple Silicon Macs and 64-bit Windows wrap the same interface, and headless runs, an Agent Client Protocol server, and a Python SDK cover scripts and embedding. There are no DeepSeek-hosted cloud sessions and no mobile app. For work that continues with the laptop closed, the same web interface runs on a VPS or home server and any browser reaches it, a phone's included, over a private network, a tunnel, or a reverse proxy; the project's safety notice asks for a disposable machine or container with the least access possible.
Code lives on GitHub by default, where an optional webhook overlay starts a read-only review session whenever a pull request is marked ready, or on GitLab. A model aggregator, a server with private access, CI/CD, Docker, and an AI code reviewer are common additions; none of them change how the harness works.
Key Features
- ✓Open-source, MIT-licensed agent whose model adapter, tools, loop, sandboxes, and interface are all replaceable plugins
- ✓DeepSeek models by default through a direct adapter, plus Kimi, GLM, Claude, GPT, or any compatible endpoint
- ✓A local web interface and desktop apps for Apple Silicon Macs and 64-bit Windows
- ✓Headless runs, an Agent Client Protocol server, and a Python SDK from the same launcher
- ✓Operating-system sandboxing on Linux, macOS, and Windows, with approval prompts under a permission policy
- ✓Subagents, skills, and MCP servers, with community plugins under the dsh-plugin GitHub topic
- ✓A GitHub webhook overlay that opens a read-only review session when a pull request is ready
When to Use DeepSeek Harness Dev Stack
- →Customizing a coding agent at the architecture level by swapping its loop, sandbox, or interface
- →Running agentic coding on DeepSeek's per-token prices with a graphical interface instead of a terminal
- →Embedding the same agent in Python tooling or automation through the SDK or the Agent Client Protocol
- →Reviewing pull requests automatically when they leave draft, from a server GitHub can reach
- →Comparing DeepSeek, Kimi, GLM, Claude, and GPT on the same tasks in one agent
Pros
- Free and open source, with no paid tier or seat fee of its own; the bill is model usage
- Every component is a replaceable plugin, so customization goes far beyond settings and hooks
- DeepSeek's models are among the cheapest capable coding models per token
- Not tied to DeepSeek: other providers and self-hosted endpoints work in the same interface
- A browser interface means a server deployment is reachable from a phone with no extra app
Cons
- Developer preview: the project warns of breaking changes and says it hasn't had a security audit
- No terminal interface ships by default, so terminal-first developers get a browser or desktop window instead
- No hosted cloud sessions or mobile app; away-from-desk work runs on hardware you keep on
- Desktop builds cover only Apple Silicon Macs and 64-bit Windows, so Linux users run the web interface
- DeepSeek's rates double during weekday peak hours, which fall in the European morning
LLM Options for DeepSeek Harness Dev Stack
DeepSeek Harness Dev Stack powered by DeepSeek
DeepSeek is the default here and the model family the harness is built for, reached through its own direct adapter with reasoning controls. V4.1-Flash costs $0.15 per million input tokens off-peak with a 1M-token context, V4-Pro takes the harder tasks at $0.66, and weekday peak hours bill at double.
DeepSeek Harness Dev Stack powered by Kimi
Moonshot's Kimi sits in the harness's bundled provider list and connects with a Moonshot API key. K3 handles the hardest agentic tasks at $3 in and $15 out per million tokens, and K2.7-Code covers everyday coding at $0.95 and $4, a step up from DeepSeek in price.
DeepSeek Harness Dev Stack powered by GLM
Z.ai's GLM is a bundled provider too, billed per token on a Z.ai API key: GLM-5.3 at $1.40 in and $4.40 out per million tokens, or GLM-5.3-Flash at $0.15 and $0.50 for quick side tasks. Its open weights can also run on your own server and join as a custom endpoint.
DeepSeek Harness Dev Stack powered by Claude
Claude connects from the bundled catalog with an Anthropic API key, billed per token: Sonnet 5.5 at $2 in and $10 out per million, Opus 5.5 at $4 and $20. The pick for the hardest multi-file changes, at many times DeepSeek's price per task.
DeepSeek Harness Dev Stack powered by OpenAI
OpenAI's GPT models connect with an OpenAI API key over the Chat Completions or Responses protocol: GPT-6.1 Sol at $2 in and $10 out per million tokens for demanding work, or GPT-6 Luna at $0.10 and $0.50 for focused, high-volume tasks.
These are highlighted picks. To see all the tools, check the LLM category.
Version Control Options for DeepSeek Harness Dev Stack
DeepSeek Harness Dev Stack with GitHub
GitHub is the default here: the harness ships an opt-in webhook overlay that opens a read-only review session in the matching workspace whenever a pull request moves from draft to ready, provided GitHub can reach the harness at a public HTTPS address. Community plugins are published under the dsh-plugin topic.
DeepSeek Harness Dev Stack with GitLab
GitLab bundles source control, merge requests, and CI/CD in one platform you can self-host. The harness works with any git remote but has no GitLab integration of its own, so automated reviews come from GitLab's pipelines or a separate AI reviewer.
DeepSeek Harness Dev Stack Add-ons
Each addition below extends this stack with a capability the base stack works fine without. None are required: include the ones your product actually needs when building this stack, and skip the rest.
Server Add-ons
Add a server when you want the agent to keep running somewhere other than your own laptop, reachable at any time and not tied to your machine staying on. Some agents also offer their own managed cloud sessions as an alternative to self-hosting; check the stack's own description for details.
DeepSeek Harness Dev Stack with Hetzner
Hetzner is the default here: CX23 has 4 GB of RAM for about €5.49/mo, enough for the harness's Node.js web server plus builds and tests, since the models run at the provider. DeepSeek Harness documents no memory minimum. A plain Linux VPS billed by the hour up to a monthly cap.
DeepSeek Harness Dev Stack with DigitalOcean
DigitalOcean's Basic Droplet with 4 GB of RAM and 2 vCPUs costs $24/mo, billed per second up to that cap, with snapshots to roll back after a risky unattended run. Its 1-Click agent images cover other tools, not DeepSeek Harness, so this is a plain Linux box for the web interface.
DeepSeek Harness Dev Stack with Hostinger
Hostinger's KVM 1 has 4 GB of RAM at $6.49/mo introductory, renewing at $11.99/mo, enough for the harness with the models served remotely. A budget pick on a plain Linux image; Hostinger's agent templates target Claude Code and OpenClaw rather than this harness.
DeepSeek Harness Dev Stack with Raspberry Pi
A Raspberry Pi 5 is a one-time hardware cost instead of a monthly bill, kept on at home to hold the harness's web interface. DeepSeek ships Linux arm64 builds of its runtime, and an 8 GB board leaves room for builds and tests next to it while the models run at the provider.
These are highlighted picks. To see all the tools, check the Hosting & Cloud category.
Remote Access Add-ons
Add remote access when you want to reach an agent running on another machine — a VPS or a home server — without exposing it to anyone but you.
DeepSeek Harness Dev Stack with Tailscale
Add Tailscale alongside a server to open the harness's web interface from a laptop or phone browser anywhere, on a private address and with no port open to the internet. The web server trusts only loopback until the tailnet hostname is named as trusted. The personal tier is free.
DeepSeek Harness Dev Stack with Cloudflare Tunnel
For a public HTTPS address without opening a port: cloudflared connects outbound from the server, matching the tunnel-or-proxy setup the harness's deployment guide describes. It's also how GitHub's webhooks reach the review overlay. Put Cloudflare Access in front, since the printed launch URL carries a credential. Free, with a domain on Cloudflare.
Code Review Add-ons
Add code review when you want an AI reading every pull request before it merges: it comments on the diff so bugs, security issues, and inconsistencies surface before a human has to catch them.
DeepSeek Harness Dev Stack with CodeRabbit
Add CodeRabbit when pull requests from the harness should get a second reviewer that isn't the same agent on the same model: it installs as a GitHub or GitLab app, summarizes the diff, and leaves line-level comments, with a free tier for open-source repos.
DeepSeek Harness Dev Stack with Greptile
Add Greptile when review should read the whole repository rather than only the diff: it indexes the codebase so its GitHub and GitLab comments carry wider context, and its self-hosted deployment keeps code on your own infrastructure.
DeepSeek Harness Dev Stack with Cursor Bugbot
Add Cursor Bugbot when a bug-focused reviewer is enough: it reviews GitHub or GitLab pull requests whatever agent wrote them, though it bills through a Cursor plan rather than a subscription of its own.
Model Aggregator Add-ons
Add a model aggregator when you want one API key and one bill for models from many providers, with automatic fallback when one of them is down, instead of setting up each provider separately.
DeepSeek Harness Dev Stack with OpenRouter
Add OpenRouter when one key and one credit balance should cover DeepSeek and hundreds of other models, with fallback when a host is down. It's a ready-made route in the harness's bundled provider catalog, and it adds 5.5% on credit purchases.
DeepSeek Harness Dev Stack with LiteLLM
Add LiteLLM when you'd rather run the gateway yourself: an MIT-licensed proxy that puts DeepSeek, Kimi, GLM, and local models behind one OpenAI-compatible endpoint, with per-key budgets and spend logs. The harness adds it as a custom model API, and self-hosting it is free.
These are highlighted picks. To see all the tools, check the AI Model Aggregators category.
CI/CD Add-ons
Add CI/CD when you want a dedicated pipeline for running tests, linting, or multi-stage builds before a deploy goes out. Many hosting platforms already redeploy automatically on every push on their own — a CI/CD tool adds the most value on top of that by gating the deploy on a passing test suite, and matters even more when the hosting choice does not auto-deploy at all, such as a self-hosted server.
DeepSeek Harness Dev Stack with GitHub Actions
Add GitHub Actions when every push should run the tests the harness writes. Its headless mode, which runs one task, prints the answer, and exits, fits a job step too, billed to a provider API key.
DeepSeek Harness Dev Stack with GitLab CI/CD
Add GitLab CI/CD when the project lives on GitLab: pipelines defined in a .gitlab-ci.yml file run the tests on every push, on GitLab's runners or entirely on your own, with the harness's headless mode as one more job if you want it.
These are highlighted picks. To see all the tools, check the CI/CD Pipelines category.
Containerization Add-ons
Add containerization when you want the app packaged the same way across local development, staging, and production, or need to deploy somewhere that isn't a managed serverless platform.
DeepSeek Harness Dev Stack with Docker
Add Docker for the isolation the harness's own safety notice asks for: a container that holds only the repository and keys one agent needs, since its sandbox reduces risk without guaranteeing it. Docker also gives the project the same setup locally, on a server, and in CI.
These are highlighted picks. To see all the tools, check the Containerization category.
Frequently Asked Questions about DeepSeek Harness Dev Stack
Can DeepSeek Harness keep working when my computer is off?
Only on another machine you keep on. There are no DeepSeek-hosted sessions and no mobile app: the web interface and the desktop apps run the agent on the computer in front of you, which has to stay awake while a task runs. For work that continues with the laptop closed, the same web interface runs on a VPS or home server, and any browser reaches it, including a phone's, over Tailscale, a Cloudflare tunnel, or a reverse proxy. Nothing extra is needed on the phone, since the interface is a web page.
Can I use Claude, GPT, or a local model instead of DeepSeek?
Yes. Anthropic, OpenAI, Kimi, and GLM are in the bundled provider list and connect with their own API keys, billed per token; signing in with a chat subscription isn't supported in the settings. A relay, a company gateway, or a model server on your own hardware joins as a custom endpoint speaking the OpenAI or Anthropic protocol. The model is chosen per session: picking one sets the default for new sessions, while a session that has already started keeps its model, so a cheap model can run routine sessions and a stronger one the hard tasks.
How does DeepSeek Harness compare to Claude Code?
Claude Code is Anthropic's agent for the terminal and IDEs, with hosted sessions on the web and Remote Control from a phone, built around Claude models on a paid plan or API key. DeepSeek Harness has neither hosted sessions nor a phone app, and it lives in a browser or desktop window rather than the terminal. What it offers instead is cost and control: DeepSeek's per-token prices run a fraction of Claude's, any provider can be swapped in, and every part of the agent, from the loop to the sandbox, is a plugin you can replace. Claude Code is the more finished product; the harness is still a developer preview.
Is DeepSeek's peak-hour pricing worth planning around?
It depends where you work. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, when every DeepSeek rate doubles; weekends and Chinese public holidays are off-peak all day. That leaves the US working day off-peak, while a European morning falls inside the second window. Even at peak, V4.1-Flash costs a fraction of Claude or GPT per token, so the doubling matters most for long unattended runs on V4-Pro, which are cheaper started in the European afternoon. The Pricing section has the per-model rates.
Is it safe to put the web interface on a public address?
Not on its own. The web server listens only on loopback, and anyone holding the printed launch URL gets in, because that URL carries a credential the session cookie then replaces. The project's safety notice adds that the harness hasn't been audited, can run model-generated commands, and should get the least access possible. Reaching it over Tailscale avoids a public address entirely. A public address through a tunnel or reverse proxy needs HTTPS and a login layer such as Cloudflare Access in front, and the server should hold only the repositories and keys that one agent needs.
Stacks Related to DeepSeek Harness Dev Stack
OpenCode Dev Stack
Developer
OpenCode in the terminal, desktop app, or browser, running open-weight models like DeepSeek, GLM, and Kimi.
Codex Dev Stack
Developer
OpenAI Codex on your desktop, in your terminal, or in the cloud, steerable from your phone.
Local LLM Dev Stack
Developer
A coding agent on a model that runs on your own machine, with no code sent to a provider.
Cline Dev Stack
Developer
Cline's open-source agent in the terminal or its desktop app, running GLM, Kimi, Claude, or any model you choose.
Scores
Popularity3/5
One of the most-starred projects on GitHub with a plugin ecosystem forming quickly, but as a recent developer preview it has far fewer working users than Claude Code, Codex, or OpenCode.
Learning Curve3/5
A first task runs minutes after launching the web interface with a DeepSeek key, but going beyond defaults means learning profiles, plugin bundles, and layered configuration, and releases still break compatibility.
Flexibility5/5
Every component from the agent loop to the interface is a replaceable plugin, any compatible provider or self-hosted model can be added, and the same launcher serves web, desktop, headless, Agent Client Protocol, and SDK use.
Performance3/5
DeepSeek's models code well for their price, but results track the chosen model, the harness is a developer preview with frequent breaking releases, and nothing runs off hardware you keep on.
Portability5/5
MIT licensed, runs on your own machines and servers, keeps sessions locally, and every model underneath can move to another provider or your own hardware.
Tools in the DeepSeek Harness Dev Stack Stack
Development Tools
Add-ons (optional — add any, or none)
Server
Remote Access
Code Review
Model Aggregator
CI/CD
Containerization
DeepSeek Harness Dev Stack Pricing
DeepSeek Harness is free and open source, so the bill is the models. DeepSeek's API starts at $0.15 per million input tokens off-peak for V4.1-Flash and doubles in weekday peak hours; Kimi, GLM, Claude, and GPT bill per token on their own keys, and a model on your own hardware costs only the hardware. GitHub or GitLab are free for individuals, and an optional server runs about €5.49–$24/month.
MIT-licensed agent with a web interface, desktop apps, headless mode, and a Python SDK; no paid tier of its own.
DeepSeek V4.1-Flash $0.15/M input and $0.60/M output off-peak, V4-Pro $0.66/$1.98, double at weekday peak; Kimi, GLM, Claude, and GPT on their own API keys at their per-token rates.
Free tier covers individual devs; GitHub Team is $4/mo, GitLab Premium is $29/mo for more seats.
OpenRouter charges provider prices plus 5.5% on credit purchases; LiteLLM is free to self-host.
Tailscale's personal tier is free; Cloudflare Tunnel is free, with Access free for up to 50 users and a domain on Cloudflare.
4 GB leaves room for builds next to the harness: about €5.49 at Hetzner, $6.49 intro at Hostinger ($11.99 renewal), $24 at DigitalOcean; or a Raspberry Pi as a one-time cost.