DeepSeek Harness: set up an agent and budget its work

DeepSeek Harness (dsh) is DeepSeek AI’s MIT-licensed agent environment. It combines model connections, workspace tools, sessions and plugins around a Cordis runtime. Use it to work on a repository or build an agent workflow whose model and tools you configure. Installing Harness on your computer does not mean inference is local: an API-backed model still sends requests to its provider.

Harness is in developer preview. Begin in a disposable project with the access needed for one task, and keep backups before upgrading or allowing changes. The project’s safety notice recommends a dedicated environment because model-generated commands and installed plugins can act on the host.

Start the Web UI and choose a workspace

DeepSeek Harness starts a local Web UI on loopback port 3080 by default. With Node.js installed, use the official npm entry point:

npx @deepseek-ai/dsh web

A local launch opens the Web UI at http://127.0.0.1:3080 by default. The server prints its address; --no-open starts it without opening a browser. For repeatable evaluations, record the exact installed package version and compare it with the release notes. The listed 0.2.1-rc.5 release is a prerelease, not a stable compatibility promise.

Open Settings → Models, enter your DeepSeek API key and save it. Next, click Choose workspace, add the directory for the trial and select it. A fresh UI requires a selected workspace before its composer becomes available. Start with a bounded reading task:

Summarize this repository and identify its main packages. Do not modify files or run commands.

Review the response, then request one change with an observable result and existing tests. Permission prompts follow the active policy; choose that policy deliberately before allowing execution. For a remote server, follow the public deployment guide rather than exposing the local service directly.

Access DeepSeek Harness from another machine

The local address refers to the machine running DeepSeek Harness. For a server reached through a public hostname, place the loopback listener behind a configured reverse proxy. The documented --public-url setting tells Harness which browser-facing address to advertise; it does not bind a public listener or authorize that hostname. --trusted-host controls the accepted authority, while the launch token and session cookie authenticate access.

For a path prefix such as /ui/, the proxy must strip that prefix before forwarding, preserve the browser-facing Host, forward WebSocket upgrades and adjust cookie scope. Keep the browser-facing connection on HTTPS and restrict direct access to the listening port. A working initial page alone does not verify session cookies or live updates. Follow the linked deployment guide for the proxy details, then check a session through the actual external address before using it for project work.

Choose an API provider or a local inference server

Use Add model provider for another provider included in the installed catalog. Custom model API accepts a self-hosted server, company gateway or another compatible endpoint. Supply the base URL, credential, model list and the protocol that endpoint actually supports: OpenAI Chat Completions, OpenAI Responses or Anthropic Messages.

A local inference server has its own checkpoint, memory and tool-format requirements. Harness is the client around it. Keep the exact endpoint and model ID with your task record, and test tool calls before giving a model a longer workflow. The GLM-5.3 guide covers one large-model option; Muse Glimmer 30B covers a smaller local-agent model.

Saved credentials are write-only in the model settings UI; the provider guide documents their storage separately from the settings references. Provider configuration takes effect on the next request without a server restart. A detected environment credential is not a connection test, so verify a small request before scheduling work.

Create a plugin with Creator mode

Creator mode is DeepSeek Harness’s workflow for agent-authored plugins and presets. In Plugins, open the Add plugin menu and choose Let the agent create a plugin. This enters Creator mode and preserves an unsent draft. It starts work after you send the request. Describe a small capability and its acceptance conditions, such as adding a panel that reads a specific local report; review the proposed package and permissions before installation.

Creator can author persistent bundles, including custom presets. Installed host plugins execute outside the workspace sandbox, and profile changes can affect other sessions using that profile. The plugin manager requires full access or approval for each management call; a workspace-only policy does not by itself contain installed host code. Keep generated plugins in a trial profile until their behavior is understood.

For existing plugins, the experimental Custom plugin markets bundle adds a Discover tab. Add a trusted HTTPS index source and inspect the package or repository/version shown before confirming installation. The plugin markets guide explains compatibility selection and installation; listing in an index does not establish who authored or reviewed the code.

Calculate model, tool and host costs separately

Harness has no fixed token tariff: the configured provider bills the model requests. Add all input, cached input and output usage across the session, including failed attempts and repairs. Search, browser services or other plugins may have separate charges. A local model replaces hosted token billing with hardware, energy and operation costs.

For a hosted DeepSeek model, use the exact model’s current API prices and the DeepSeek-V4.1-Flash guide to understand the billing categories. A provider key or coding subscription used in one integration does not establish billing terms for another.

Track the host separately: an existing laptop has incremental energy and operator time; a rented machine has reserved runtime and storage charges. With an assumed $0.50/hour host, a 30-minute session adds $0.25 before model and tool usage. Divide the complete cost by results accepted after checking them. Plugin generation, installation failures and retries can consume time or tokens without producing an accepted task.

Keep a session record for cost comparisons

Record the DeepSeek Harness package version, provider and exact model ID, chosen preset, workspace revision and enabled plugin versions before a trial. Save the task’s acceptance conditions as well: a repository summary, a patch passing existing checks and a deployed result are different outcomes. This makes it possible to compare runs without attributing a changed tool configuration or larger task to the model alone.

After the session, collect billed usage across every attempt, elapsed host time, paid tool calls and the number of retries. Note which results passed review and what still needed manual repair. Keep request usage separate from the final answer’s visible length, since reasoning and repeated context can contribute to billing. Save a redacted record with the accepted artifact; exclude API keys, launch tokens and private source content from shared cost reports.

Diagnose version-specific setup problems

These reports concern older DeepSeek Harness prereleases. Keep their package version and client profile with any troubleshooting steps.

JounQin’s October 5 report traces a startup failure in older prereleases to peer dependencies skipped with npm’s legacy-peer-deps setting. Its update says installing just the first missing package was insufficient. Read the full startup diagnostics and package-version context before applying an old workaround to a newer installation.

neuviletteflux reported a Creator-mode selection reverting to Standard on the 0.2.1-rc.2 macOS desktop profile, while a Web profile of the same package worked. This October 21 report concerns rc.2, not the separately listed rc.5 prerelease. The report does not establish whether rc.5 fixes that behavior. Verify the selected mode before sending a task and compare the current release’s behavior; it does not establish that all Creator sessions fail.

When to add dsh-superagent

Start with Harness for manually initiated sessions, model setup and plugins. dsh-superagent adds a separate research prototype that watches repository requests and routes them into ticketed coding work. Its account, permission and always-on operating requirements go beyond this first-session setup. You can use Harness without that watcher.

Sources

Reviewed November 3, 2026: Harness README, Web quickstart, provider setup, safety notice, preset UI, plugin manager, Custom plugin markets, release notes and the two linked community reports. No fixed per-task cost is inferred from the user reports.

For native desktop interaction, compare agent-desktop. Find related workflows in the Agent guides.