MCP-first coding agent for digital studios.
Powering creative work across Blender, game engines, and production tools.

_中文文档

---

See it in action

Web UI — streaming chat, tool calls, permission dialogs

Terminal UI — real-time streaming, slash commands, inline rendering

Tip: GIFs auto-play in GitHub. Click any image for full-size view.

---

What makes dscode different

🔌 MCP-First Digital studios don't use one tool. They use ten. MCP turns every tool into an API — dscode is the agent that orchestrates them. Blender for 3D modeling, PlayCanvas for real-time graphics, browser automation for testing, documents for specs, spreadsheets for data. If your production tool has an MCP server, dscode brings it into the workflow. Your toolchain. One agent. All through MCP.

🧬 Spec-Driven Development dscode is built entirely through spec coding with OpenSpec. Every feature begins as a formal spec — openspec/specs/ is the source of truth, code is the implementation. We don't encourage manual commits; all design and development flows through the SDD pipeline. Code is the implementation of specs — not the other way around.

🔍 MCP Tool Search Too many MCP servers? Context explosion is a real problem when every tool schema competes for token budget. dscode ships with a built-in search_tools driver — MCP tools are discovered on-demand by the model, not pre-loaded. Only the tools actually needed enter the context window. Connect dozens of MCP servers without worrying about overhead. All the tools. None of the bloat.

🐋 DeepSeek V4 Pro — Recommended dscode is purpose-built for DeepSeek V4 Pro — our recommended model for digital creation. Its reasoning depth handles complex multi-tool workflows ("model this in Blender, render in PlayCanvas, document the result") without losing context. A vision model fallback pipeline transparently routes screenshots and reference images to vision-capable models. Prompt caching is tuned to maintain 97–99% cache hit rates via prompt_cache_key affinity and prefix-stable message construction. Every optimization is measured against DeepSeek's API behavior. DeepSeek V4 Pro. For when your toolchain needs more than autocomplete.

---

Capabilities

🖥 Terminal + Web Full TUI with streaming, thinking, tool calls. Modern React Web UI with identical feature parity via WebSocket.	🔌 MCP Connector Stdio, Streamable HTTP (MCP 2025-11-25), legacy SSE fallback. Auto transport inference. MCP App sandbox for server-driven UI.	🛡 Agent Harness Permission control, context compression (1M token window), session persistence, cross-session memory, retry with exponential backoff.
📦 Skills System Declarative third-party extensions via SKILL.md. On-demand activation. User-level + project-level scopes.	👁 Vision Pipeline Auto-routing to vision-capable models. tesseract OCR fallback (ENG + CHI). Drag, paste, or @-file images.	🔧 Built-in Drivers read_file, write_file, edit (hash-anchor), bash, grep, glob. MCP tools discovered on-demand via search_tools.

---

MCP in 30 seconds

// ~/.dscode/settings.json
{
  "mcpServers": {
    "blender": {
      "command": "uvx",
      "args": ["blender-mcp"]
    },
    "playwright": {
      "command": "npx",
      "args": ["@anthropic/mcp-playwright"]
    }
  }
}

dscode auto-connects on launch. Tools appear as mcp_blender_* and mcp_playwright_*. MCP servers can also serve sandboxed UI via the App Host — no boilerplate, no SDK, no glue code.

See what MCP can do

PlayCanvas + MCP — build a jump game entirely through natural language

Blender + MCP — 3D modeling and scene composition through conversation

Tip: These are real MCP workflows — dscode orchestrates PlayCanvas and Blender as if they were native APIs.

---

Install

npm install -g @wangcan26/dscode
dscode              # Terminal UI
dscode --web        # Web UI → http://localhost:3000

First launch? Run /config key <your-api-key> and /config model deepseek-v4-pro to get started. Type /help for the full guide.

Build from source:

git clone https://github.com/wangcan26/dscode.git
cd dscode && npm install && npm run build
node dist/dscode.mjs

---

Harness Philosophy

We follow Occam's razor in harness design. dscode does not pre-build intent understanding modules, plan modes, or elaborate agentic scaffolding until the system prompt proves insufficient. Most coding agents pile on pre-turn planning, reflection loops, and multi-agent orchestration upfront — we wait until the model demands it.

That doesn't mean the harness is bare. It means every piece earns its place.

One example where we went deeper: the edit tool. Based on @_can1357's hash-anchor protocol, our edit tool replaces fragile line-number and regex-based editing with a content-addressable anchor system (spec):

• Three-level adaptive resolution — ambiguous 6-char hashes are resolved silently through 8-char → context-augmented (3-line window) matching before rejection
• Occurrence + line-hint disambiguation — occurrence: 3 picks the Nth match; line field auto-selects the closest candidate, rejecting only when equidistant
• Proximity-based range resolution — for range endpoints, if one side is unique the other automatically resolves to the nearest candidate in the correct direction
• Low-entropy filtering — lines like } are rejected as anchors; the tool returns up to 6 neighboring [high] anchors as alternatives
• Atomic batch + overlap detection — 6 operation types in one call, all-or-nothing. Overlapping ranges within a batch are detected and rejected
• Checkpoint + safety rollback — file checkpointed before edit. Post-edit sanity checks (duplicate lines, delimiter balance, orphan else) roll back suspicious changes
• Structured invalidation scope — anchors_valid_through and must_refresh_from_line tell the model exactly which anchors survive, enabling chained edits without re-reading
• Localized diff with live anchors — successful edits return a diff with fresh 6-char hashes, so the model can continue editing immediately

dscode builds dscode. This edit tool — combined with our spec-driven workflow — is what enabled dscode to develop itself. Every feature, from the hash-anchor protocol to the checkpoint system, was implemented by dscode running on DeepSeek V4 Pro, editing its own source tree through MCP-driven tools. It's not a demo. It's how this project ships.

This is the kind of harness work we invest in: not adding more AI, but making the AI's tools dependable.

---

Learn more

Document	What's inside
ARCHITECTURE.md	Full architecture: Agent as OS, 6-layer design, Driver/Skill model, source tree
ROADMAP.md	What's next: Sub-Agent system, System Prompt modularization, Diff-based editing
CONTRIBUTING.md	How to contribute: philosophy alignment, OpenSpec SDD workflow, coding conventions
STYLE.md	TypeScript coding style: naming, imports, module structure, error handling

---

Acknowledgments

dscode stands on the shoulders of:

• OpenSpec — the spec-driven development framework that shapes our entire workflow

• pi-ai / pi-agent-core — Mario Zechner's agent loop and model abstraction foundation

• taste-skill — Leonxlnx's design taste skill system, inspired our skills architecture

• @_can1357 — hash-anchor editing protocol, the cornerstone of our edit tool

---

_{If you like this project, give it a ⭐ Star — your support keeps dscode evolving.}