DeepSeek Harness Can Write. It Can't See, Hear, Create, or Publish. AnyCap Can.

Text was never the missing piece

Someone asked us whether DeepSeek Harness can write an article about AnyCap. Sure — any model writes text without help. That's a training-data question, not a harness question.

The real question is what DeepSeek Harness can't do without help: generate an image, cut a video clip, pull a fact from this week's web instead of last year's training run, or hand back a finished page instead of a local file only the operator can open. DeepSeek frames its own project as "Agent = Model + Harness" — the model reasons, the harness does everything else. AnyCap is built on the same seam from the other side: agents reason, plan, and write code just fine, then stop the moment you ask for an image.

What DeepSeek Harness ships

DeepSeek Harness (CLI name dsh) is DeepSeek's open-source agent runtime, built on Cordis, a plugin kernel where models, tools, skills, sessions, sandboxes, and storage are all swappable modules. Nothing is hardcoded — capabilities mount and unmount through config.

Four ways to run it

• Standard mode — the full toolset: file editing, shell, web search, skills, planning, subagents.
• Code mode — the same tools exposed through an SDK, so the model chains multi-step operations in one program.
• Minimal mode — just a shell and a file editor, for benchmarking a model in isolation.
• Creator mode — for building and testing new presets.

Protocols it already speaks

DSH ships an MCP client, Agent Client Protocol support, and reads AGENTS.md / CLAUDE.md files, the same conventions Claude Code, Codex, and Cursor already use. Every tool call, reasoning step, and subagent handoff lands in an append-only session log that can be resumed, forked, or replayed.

DSH isn't trying to do everything itself. It's trying to be a well-logged place to plug things into.

What AnyCap ships — the part DSH leaves out

AnyCap puts its own pitch bluntly on its homepage: agents reason and plan, then stop at "I can't do that." AnyCap is the last mile between a plan and a finished, deliverable output, and it groups that last mile into five verbs.

Create

Generate images, video, and music across 20+ models behind one API and one bill — Seedance, Kling, Suno, GPT Image, Nano Banana Pro, and others.

Understand

Read a screenshot, review a screen recording, transcribe a customer interview.

Search

Grounded web search with citations, or crawl a specific competitor page instead of guessing what's on it.

Manage

Store every generated asset in one place instead of scattering files across a session.

Deliver

Publish the result as a live page or a shareable link — not a file stranded on whatever machine ran the agent.

Wiring AnyCap into DeepSeek Harness

The install is the same regardless of which agent ends up calling it:

curl -fsSL https://anycap.ai/install.sh | sh
anycap login
anycap status

Option A — Agent Skill

npx -y skills add anycap-ai/anycap -s '*' -g -y

DSH's Standard mode already reads Skills natively. Drop this in and the model routes plain-language requests to the right AnyCap workflow on its own.

Option B — MCP server

{
  "mcpServers": {
    "anycap": {
      "type": "stdio",
      "command": "anycap",
      "args": ["mcp", "--allow-root", "/absolute/path/to/project"]
    }
  }
}

DSH ships an MCP client out of the box, so this is the same registration flow used for any other MCP server. No custom adapter to write.

Nothing stops you running both at once — the Skill for routing, MCP for typed execution.

What it looks like end to end

Here's an actual DeepSeek Harness session, asked in plain language to install AnyCap, walked through step by step.

Step 1 — Ask the agent, and it finds the Skill on its own

The user types "help me install anycap." DSH reads its Skill catalog, matches anycap-cli, checks the binary, finds AnyCap already installed but not yet authenticated, and starts the device login flow — no manual setup instructions needed.

Step 2 — Approve the device login, the agent verifies it

The agent hands back a one-time code and a URL. You authorize in the browser; it polls the session on its own and confirms: "AnyCap is installed and authenticated!"

Step 3 — The agent confirms every capability is live

A final anycap status check lists every active capability — image, video, music, and audio generation, media understanding, web search and crawl, Drive, Pages — the same five verbs from the section above, now live inside a DeepSeek Harness session.

Why the pairing isn't a stretch

• Both projects made the same bet on "capability" as the unit of composition, not "feature." DSH's plugin kernel defines how a runtime accepts new abilities; AnyCap's Skill/MCP/CLI trio defines a bundle of abilities that fits any runtime built that way.
• DSH is open source and keeps a full session log, which matters once an agent is generating and publishing real content — every image and every claim traces back to the tool call that produced it.
• AnyCap doesn't care which agent is calling it. It authenticates once and exposes the same tools to Claude Code, Codex, Cursor, or DeepSeek Harness. Plugging into a newly open-sourced harness isn't a special case for AnyCap. It's the default one.

The takeaway

DeepSeek Harness can already write about AnyCap on its own — that was never in question. What it can't do on its own is generate the illustration at the top of this article, check a claim against this week's web, or publish the result as a page instead of a local file. Add AnyCap, and it can. DeepSeek Harness gets the agent thinking and planning. AnyCap lets it finish what it started.