AI News

DeepSeek Harness and Creator Mode: Plugins, Pricing, Specs and Benchmarks

Updated November 2, 2026. Originally published August 17, 2026.

DeepSeek Harness can turn a request into a new tool inside the agent application: a panel, an editor, a visualization or another plugin. Creator Mode is the route for asking the agent to build those additions. The useful starting point is a small tool with a clear job and an output you can check. A convincing interface alone tells you little about whether its calculations, saved files or exports work.

DeepSeek’s official Creator Mode collection shows seven examples and acknowledges that some required several rounds of refinement. Treat them as ideas for what to build.

Kingy reviewed public documentation, source code declarations and published research for this update. We did not install Harness, reproduce the demonstrations, run model calls or test its security controls. Practical prompts and acceptance checks below are our proposed workflow, not reported hands-on results.

DeepSeek Harness open-source AI agent

What Harness, Creator Mode and the model each do

DeepSeek Harness, usually abbreviated DSH, is the application and runtime around an AI model. Its plugins can supply tools, skills and interface components. DeepSeek presents it as a public preview for coding, documents, data analysis, research and background work. The source is MIT-licensed, and the repository warns that preview interfaces can change. Product overview, official repository.

The distinction becomes clearer with a simple example. Suppose you want a sidebar that groups a list of links by reading status. The model interprets the request and writes code. Harness supplies the application where that code runs. A plugin supplies the sidebar, controls and storage behavior. Creator Mode organizes the work of creating that extension. None of those labels guarantees that links will survive a restart; you still need to test that requirement.

At the developer level, a basic plugin is a TypeScript module with an apply(ctx) entry point that registers capabilities through its context. Users can ask the agent to produce that code, while developers can inspect and maintain it directly. First-plugin guide.

The underlying Cordis research describes a plugin runtime with dependency management, hot replacement and cleanup of registered effects. That architecture helps explain how the application can change while running. Cleanup of an interface component does not undo a file it overwrote or a request it sent to an external service. Cordis paper.

Creator Mode therefore has no separate parameter count, training dataset or context window. Those specifications belong to the selected model. Rendering speed and export support also depend on the generated plugin and its dependencies.

Availability, versions and system requirements

The installation route matters. A desktop installer, an npm release, current source code and the Python SDK can have different requirements.

Route Documented position checked November 2
Official desktop Apple silicon Mac, with macOS 13 or later; Windows x64
Browser interface Local Web UI, normally at http://127.0.0.1:3080
npm latest and next 0.2.1-rc.5
npm alpha 0.2.2-alpha.2
Node for 0.2.1-rc.5 ^22.19.0 || >=24.0.0
Node in checked current source >=24.9.0
Python SDK Python 3.10+; Linux x64/arm64, macOS 14+ arm64 and Windows x64

Sources: downloads, repository, npm registry, rc.5 manifest, checked source manifest, SDK guide.

The download page does not list a Linux or Intel Mac desktop installer. Linux SDK support is a different claim. Packaged desktop software bundles its runtime; the Node requirement is relevant when using the npm/source route. Node 24.9 or later satisfies both manifests checked here.

The rc.5 npm package was published October 29 UTC, followed by its GitHub release on October 30 UTC. These are separate distribution records, and neither establishes the version installed on your computer. Recent release notes include plugin creation entry points, Flash 1017 support and Search/Fetch support. Features documented on the alpha source branch should not automatically be attributed to every desktop build. Registry metadata, release notes.

No universal minimum RAM, GPU, storage requirement, Creator task quota or plugin export specification was established in the reviewed material. A hosted model runs inference at its provider; local rendering, large files and any locally hosted model have their own hardware demands.

The seven official Creator Mode examples

The official collection contains these requested/showcased functions. It does not establish their reliability on arbitrary inputs.

Example Requested result
Landmark gallery A sidebar with 3D views of landmarks including the Eiffel Tower, Great Wall and Forbidden City
Transformer explainer An interactive 3D teaching game with carefully positioned controls
Interface styles Five application appearances, including a pixel style
Poster workbench An editable poster tool with text and shader-color controls
Video workbench Clips, preview, timeline, editing, export and conversational changes
Paper reader A sidebar reader with layout retention, selection, highlights, notes and chat questions
Conversation statistics Sessions, usage, estimated spending, budget, timeline and CPU activity

The original prompts are linked above. Their descriptions are a starting brief, not verified codec lists, financial calculations or export guarantees.

What a useful version should pass

The following are Kingy’s proposed acceptance checks. We have not run them against the showcased plugins. Their purpose is to turn a visual idea into a deliverable you can accept or reject.

3D viewers and teaching tools

For a 3D viewer, check orbit, zoom, reset and labels before adding more objects. Try a small window and keyboard navigation. Record whether interaction remains usable with the intended scene size. If historical or architectural accuracy matters, compare geometry and proportions against an identified reference; recognizable silhouettes are insufficient.

Keep asset provenance separate from appearance. Ask where models, textures and fonts came from and whether you can redistribute them. If you need a geometry file, specify its format and reopen the exported file in another application. A scene displayed on screen does not establish export capability.

For an educational tool, write down the exact concept it must explain. An attention demonstration could use a tiny numerical example with visible input dimensions and intermediate values. Check that displayed weights and outputs match the chosen example. Identify the architecture before accepting claims about normalization or residual connections. Pause, step and reset controls make errors easier to find than an uninterrupted animation does.

Themes and poster tools

A theme affects ordinary work throughout the app. Inspect menus, dialogs, selected text, code blocks and error messages in both light and dark modes. Check visible keyboard focus and readable contrast. Then disable the theme and confirm that its styles disappear. A screenshot of the main conversation view covers only a small part of that test.

For a poster tool, define the canvas size, editable fields, export format and intended viewing size. Save the project, close it, reopen it and change one line of text. Inspect the export for clipped characters, substituted fonts and transparent areas. If repeatable output matters, save the inputs and settings alongside the exported file.

A rendered design interface also leaves an unanswered product question: where do its images come from? Establish whether the plugin draws graphics, imports existing assets or calls a separate generation service, and account for the latter’s credentials and charges.

Video exports and document readers

Start a video test with a disposable short clip whose duration and audio are familiar. Trim a known interval, export it, and reopen the result in an independent player. Check duration, frame dimensions, audio synchronization and the final frame. Test cancellation and an unsupported file before trusting the tool with a long edit. Specify required formats explicitly instead of assuming a complete professional-editor feature set.

For a document reader, use a small mixed test set: a straightforward text PDF, a two-column paper, a page with equations or tables, and a scan. Check reading order and selection boundaries. Add a highlight, restart the tool and verify that it remains attached to the correct passage. Answers should identify the page and supporting text, so a reader can distinguish an extraction problem from an unsupported answer. Require a clear limitation message when OCR is unavailable.

Usage and cost dashboards

A cost panel needs traceable inputs. Keep a small fixture with known session IDs, token counts and expected totals. Separate cached input, uncached input and output; store the applicable price version and timezone. Feed it a duplicate event and a missing field. The panel should reveal those conditions instead of silently counting twice or treating unknown values as zero.

Reconcile its estimate against provider usage records before relying on it for budgeting. Local CPU activity is a host measurement; it cannot tell you the GPU work performed by a remote model service. Label measured values, calculated estimates and unavailable data distinctly.

How to start and build a first plugin

For desktop use, begin with the official download page. For the browser route, the repository documents:

npx @deepseek-ai/dsh web

An unversioned npm command follows the current latest tag. Check its requirements before running it. We have verified the documented command, not executed the installation for this article. Repository setup.

The Web quickstart directs users to configure a model in Settings → Models and select a workspace before composing. Use a disposable folder for the first attempt and inspect the active permissions. Choose Creator Mode for the plugin task; exact menu placement may differ across preview builds.

DSH supports catalog providers including Anthropic, OpenAI, Moonshot and Z.ai, plus compatible custom endpoints using OpenAI Chat Completions, OpenAI Responses or Anthropic Messages. Compatibility settings matter: accepting the protocol does not establish identical reasoning, tools or output-limit behavior. Credentials are stored locally according to the provider guide. That storage statement does not mean model inputs stay local. Model configuration.

Give the agent one functional target. Describe the inputs it may read, where the interface belongs, what it may write and how you will check success. Ask it to inspect the installed plugin interfaces before coding. The checked Creator instructions describe building a source bundle in the workspace and installing it through the plugin manager. Installation targets a profile, can affect other sessions using that profile and survives restarts. Creator development instructions.

The UI reference describes registration into supported interface locations and cleanup of styles, listeners and timers. The verification reference distinguishes a registered component from a UI that has actually rendered and been exercised. Those are useful distinctions when reading an agent’s completion message. UI reference, verification reference.

Read the installation result and warnings. Check the visible control yourself, perform its main action, then restart and repeat. Keep a copy of the generated source and your test inputs. Before expanding the plugin, verify that you can disable or remove it through the supported plugin-management path. The manager documents application results and cases where hot reload cannot fully apply a change. Plugin manager.

An original starter prompt

This Kingy-authored example is deliberately small and has not been tested. It creates a reading queue, not a full document reader:

Build a reading-queue plugin for this workspace. Put its controls in an appropriate sidebar location after inspecting the installed UI interfaces.

Store entries in reading-queue.json inside the selected workspace. Each entry has a title, URL and status: unread, reading or finished. Let me add an entry, edit its title, change its status and filter the list. Do not open URLs or send the list to an external service.

Start with the smallest functional version. Show a useful empty state. If the saved JSON is malformed, report the problem and leave the original file untouched. Confirm before replacing an existing file during setup.

Acceptance checks: three sample entries can be added; each filter shows the expected entries; changes survive an application restart; malformed JSON produces an error without data loss; disabling the plugin removes its controls without deleting the saved queue. Use disposable test data.

Tell me which checks you actually performed, what evidence you observed and what remains untested. Ask before accessing files outside this workspace or using a paid service.

Treat those instructions as requirements to verify. Text in a prompt does not enforce a filesystem boundary or prevent a generated implementation from making a mistake.

When requesting a correction, describe the observed failure precisely: “After reopening, the finished entry appears as unread; preserve the saved status and rerun the restart check.” That gives the next iteration a reproducible target. Asking for a broadly “better” plugin makes progress harder to judge.

Model specifications and pricing

These are DeepSeek’s documented API limits. DSH settings may be lower.

API model Current version Context / maximum output Vision Concurrency
deepseek-flash V4.1-Flash-1017 1M / 384K tokens Yes 2,500
deepseek-v4-pro V4-Pro-0813 1M / 384K tokens No 500

Both list thinking and non-thinking modes, tool calls, JSON output, Responses API and Anthropic API support.

USD per million tokens, checked November 2:

Model Cached input off-peak / peak Uncached input off-peak / peak Output off-peak / peak
Flash $0.003 / $0.006 $0.15 / $0.30 $0.60 / $1.20
Pro $0.022 / $0.044 $0.66 / $1.32 $1.98 / $3.96

Peak: weekdays 01:00–04:00 and 06:00–10:00 UTC, except Chinese public holidays. All other times are off-peak. Official pricing.

What those rates mean for a budget

Kingy’s hypothetical calculations, not measured plugin costs:

Hypothetical usage Off-peak cost Peak cost
Flash: 1M uncached input + 100K output $0.21 $0.42
Pro: 1M uncached input + 100K output $0.858 $1.716
Flash: 900K cached input + 100K uncached input + 100K output $0.0777 $0.1554

The last off-peak calculation is 0.9 × $0.003 + 0.1 × $0.15 + 0.1 × $0.60. The assumed 90% cache-hit share is not guaranteed.

Search and Fetch each cost $0.75 per 1,000 successful requests. The published effective date for server-side tool billing through Harness and other supported agents was November 1 at 00:00 Beijing time, or October 31 at 16:00 UTC. That date has passed. Failed requests, empty searches and timeouts are excluded under the documented rules; model-token charges remain separate. Search/Fetch billing guide.

At those rates, 100 successful searches plus 100 successful fetches add $0.15. A thousand of each adds $1.50. We have not inspected an actual bill.

A session may make many model requests while reading, editing, correcting and summarizing its work. Budget from recorded usage across the complete session. Track the human time spent fixing the result as well: a cheap failed export still leaves the job unfinished.

Free software, account access and API billing

The open-source license and inference charges answer different questions. DeepSeek’s terms distinguish signed-in official-model access from custom-model use; the latter can work without a Harness login and leaves provider charges with the user. Harness terms.

The account-provider documentation describes an account-specific model catalog and minimum client versions. We did not verify a universal free allowance, renewal interval or separate Creator subscription price. Read the offer shown in your account, and do not apply the API table above to it without checking its terms. Account-provider reference.

A plugin can also introduce another bill by calling an image service, transcription API, hosting provider or other paid tool. Include those services explicitly in its requirements and budget.

What the benchmarks establish

There are three numerical evidence sets worth separating from the demonstrations. None gives a verified success rate for recreating the current seven plugins.

Historical Creator Mode software-task evaluation

The frozen FrontierHarness v1 benchmark used Kimi K3 through Fireworks, 30 software tasks and 12 configurations across nine agent systems. Its task mix was 21 Terminal-Bench 2.1 tasks and nine DeepSWE 1.1 tasks, with one selected valid first attempt per configuration/task cell. Benchmark definition.

The August 22 snapshot reports the following. Its cost figures omit some failed-run costs; they are incomplete cost-per-pass aggregates, not plugin prices.

DSH configuration Passed Pass rate Published cost/pass, incomplete records Median runtime
Creator 19/30 63.3% $3.28 6m44s
Standard 18/30 60.0% $3.46 6m17s
PTC 18/30 60.0% $4.58 7m44s
Minimal 17/30 56.7% $4.72 5m41s

Source: published result JSON. Summing available per-task costs and dividing by passes reproduces effective_cost_per_pass. Those cost entries cover 27/30 runs for Creator, Standard and PTC, and 26/30 for Minimal. Unpriced failures still count in the pass-rate denominator.

All four DSH configurations used 0.1.0-rc.8. Runta supplied runtime infrastructure and sponsorship. Separately, the repository says its normalized fields depend on unpublished cache-adjustment data. Those fields cannot be independently rebuilt from the public inputs. Evaluation README.

Creator’s advantage over Standard was one task in this sample. The tasks were software challenges, not tests of current plugin creation, and the backend was Kimi K3 rather than Flash 1017. These results support a narrow historical comparison. They do not establish that Creator is generally superior or predict the cost of your next plugin.

The broader model–agent comparison

The October 1 preprint Finding the Right Fit studied 66 configurations across three task collections. DSH was 0.1.1-rc.2. These selected results used GPT-6 Astra on the same 63-task non-H100 Terminal-Bench 4 subset. All requested high reasoning effort through OpenRouter, pinned to first-party providers with fallbacks disabled:

Agent system Score API cost per task
Pi 60.32% $4.66
Codex 55.56% $4.61
DeepSeek Harness 52.38% $19.94
OpenHands 49.21% $5.40

Costs are the study’s OpenRouter-priced totals divided by 63 and rounded, not current direct DeepSeek prices. The same effort label need not mean equal thinking budgets. The study retained one final run per task, used differing native tools/settings and did not measure run-to-run variance. Its subsets are not official leaderboard replications. Other model pairings changed the ranking; this predates Flash 1017 and is not a current Creator-specific score. Study and limitations.

For a purchasing decision, compare the complete workflow you would use. To investigate the agent software itself, hold the model and endpoint constant where possible. Those are different experiments and can produce different answers.

DeepSeek’s backend-model scores

DeepSeek’s Flash 1017 model card reports these results:

Benchmark Previous Flash Flash 1017 Evaluation software
Terminal-Bench 4 31.2 42.3 Claude Code
DeepSWE 1.1 74.2 77.0 Mini-SWE-V2

These are vendor-reported model evaluations using the named software, not Creator Mode plugin scores. Flash 1017 model card.

The October 17 announcement says Flash 1017 was optimized for Harness and other applications. Claims of shorter or faster work remain vendor claims here. A change to the backend may help an agent, but its score in another evaluation setup cannot be transferred directly to Creator Mode.

Security and privacy details that affect real use

DeepSeek’s safety notice calls the software experimental and unaudited. Generated commands and plugins can affect accessible files, processes, credentials and networks. Permission prompts are useful controls, but they are not evidence that every execution route is isolated.

Start with copied data and limited access. Keep your production credentials outside the first test environment. Before allowing a plugin to modify valuable files, inspect where it reads and writes, how it handles errors and whether an interrupted operation leaves a recoverable state.

Several documented distinctions deserve attention:

  • Ordinary model traffic: a local application can send prompts and context to a hosted endpoint. Offline use requires a local model route and local availability of every dependency and tool the workflow needs.
  • Official-service data handling: the September 20 privacy policy describes session data that can include inputs, uploads, outputs, tool invocations and plugin configuration, with training and improvement among its stated uses. For custom models, inputs and outputs go directly to the configured provider under that provider’s policy. Third-party plugins and services introduce their own processing. Harness privacy policy.
  • Additional session logs: the checked source documents an enabled-by-default contributor for official DeepSeek model requests, with an 8 MiB default field limit. The Web setting is Settings → General → Upload Session Log when using the official model API. Re-enabling it resumes the unaccepted backlog, including events recorded while disabled. Feedback telemetry has a separate setting. Turning this contribution off does not stop normal model inputs from being sent. Session-log documentation.
  • Desktop interaction analytics: the checked source separately documents enabled-by-default interaction telemetry without a user-facing control. It excludes prompts, responses and API keys from those events; ordinary Web clients do not send those particular events. This is a source-snapshot finding, not verified traffic from a shipped installer. Analytics source.
  • Installed plugin code: the checked plugin-manager reference says installed Host code runs in the application process outside the workspace sandbox. A workspace selection therefore does not confine that code. This is another reason to inspect a plugin before approving installation. Plugin-manager reference.
  • SDK permissions: the documented sdk-minimal profile pins danger-full-access; its shell can modify paths visible to its runtime. Its working directory is not a security boundary. That is an SDK-profile detail, not a statement that Creator Mode uses the same default. SDK guide.

There is also historical external security research. Tencent Zhuque Lab’s August preprint reports 14,560 controlled runs against an August 13 DSH revision, using a deepseek-v4-flash backend through a local proxy. Its deterministic evaluator counted full attack success in 819 runs, or 5.6%; its semantic evaluator counted 772, or 5.3%. Sensitive tools were simulated, and the tested configuration lacked A.I.G prompt hardening. Those figures do not measure today’s desktop vulnerability rate. Security study.

Choosing a use case and fixing common failures

Our editorial recommendation is to start with a bounded tool you will use repeatedly: a small queue, a workspace-specific checker or a panel that displays data you already understand. The benefit is easier to judge when you can specify its correct behavior in a few sentences.

Keep an existing editor for work whose value depends on mature import/export support until the generated alternative passes your files. Choose an ordinary agent conversation when the task needs a one-time answer; maintaining a plugin adds little value unless persistence or repeated interaction helps. Choose the SDK when you need to operate the surrounding application yourself and can manage its configuration and execution environment. These are task-fit judgments, not new head-to-head test results.

Symptom First useful check
The agent says it installed a plugin, but nothing appears Check the active profile, installation application result, warnings and intended interface location. Ask for observed rendering evidence.
It works until restart Repeat with a tiny fixture; inspect what was saved, where it was saved and how it is loaded.
Export appears successful but the file is unusable Inspect the actual output path and file, then reopen it independently. Record the requested format and error.
Model requests fail Separate credentials, account entitlement, API balance, model compatibility and client-version errors.
A cost total looks wrong Reconcile a few raw events manually, checking duplicates, cache categories, timezone and price date.
Removing the UI leaves changed data Restore from the data backup; plugin removal should not be assumed to reverse earlier file or service changes.

For your first evaluation, save the exact app version, model, profile, prompt and fixture. Record the revisions required, actual usage charges and remaining failures. Keep the plugin only after its main action, saved state, error handling and disable/remove behavior pass the checks you defined.

Related reading: DeepSeek V4.1 Flash 1017 model guide and DeepSeek Super-agent architecture analysis.