convez.ai
MCP ServerAutomaçãoAvançado
creativedswork

dscode — Agente de Codificação para Estúdios Digitais

creativedswork/dscode

Agente MCP que conecta toda sua cadeia de ferramentas criativas em um único fluxo de trabalho

Ver no GitHub
BomPluginScore · de 0 a 100

Servidor MCP especializado em orquestrar ferramentas de criação digital (Blender, game engines, automação web) através de um agente de IA unificado. Implementa busca dinâmica de ferramentas MCP para evitar explosão de contexto e é otimizado nativamente para DeepSeek V4 Pro com caching de prompts em alta taxa de acerto.

Instalação

$npx convez-ai creativedswork/dscode

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
11
Forks
3
Acompanhando
0
Saúde
50%
Último commit
há 2 meses
Versão
v0.3.2
Issues abertas
1
Contribuidores
2

Casos de uso

  • Automação de pipelines de produção 3D (modelagem em Blender → render em PlayCanvas → documentação)
  • Orquestração multi-ferramenta para estúdios digitais sem perder contexto do modelo
  • Desenvolvimento spec-driven integrado com descoberta dinâmica de ferramentas MCP sob demanda

Veredito

Ferramenta sólida e bem arquitetada para estúdios que precisam coreografar múltiplas ferramentas criativas. A abordagem spec-driven e o caching inteligente de prompts demonstram maturidade, mas o público é nicho (produção digital avançada). Recomendado para times que já trabalham com MCP e DeepSeek.

README

A tradução para o português deste README ainda está na fila.


See it in action

Tip: Videos play inline — click to watch demos directly in GitHub.

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

Tip: In TUI, paste clipboard images with Ctrl+V (macOS) or /image clipboard.


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

Tip: Videos play inline — these are real MCP workflows, click to watch.

Install

npm install -g @creative-dswork/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/creativedswork/dscode.git
cd dscode && npm install && npm run build
node dist/dscode.mjs

Configuration

dscode uses two levels of settings.json, merged with project settings overriding user settings:

ScopePathPurpose
User~/.dscode/settings.jsonDefaults across all projects
Project.dscode/settings.jsonPer-project overrides

Note: Model configuration (provider, modelId, apiKey, thinkingLevel) lives in ~/.dscode/config.json, managed via /config commands. Type /help in-session for the full command list.

Quick reference

// ~/.dscode/settings.json
{
  // --- MCP Servers ---
  "mcpServers": {
    "blender": {
      "command": "uvx",
      "args": ["blender-mcp"]
    },
    "playwright": {
      "command": "npx",
      "args": ["@anthropic/mcp-playwright"]
    },
    "my-api": {
      "url": "https://my-mcp.example.com/mcp",
      "headers": { "Authorization": "Bearer <token>" }
    }
  },

  // --- Permissions ---
  "permissions": {
    "allow": [
      "Bash(git add *)",
      "Bash(npm *)"
    ],
    "deny": [
      "Bash(rm -rf *)"
    ],
    "rules": [
      { "tool": "Bash(curl *)", "decision": "allow", "priority": 5 }
    ]
  },

  // --- Skills ---
  "skills": ["brandkit", "minimalist-ui"],

  // --- Retry ---
  // Controls how dscode retries failed API calls (rate limits, timeouts, server errors).
  // Uses exponential backoff: starts at baseDelayMs, doubles each retry, capped at maxDelayMs.
  "retry": {
    "maxRetries": 3,           // Max retry attempts before giving up
    "baseDelayMs": 1000,       // Initial delay before first retry (ms)
    "maxDelayMs": 30000,       // Upper bound on backoff delay (ms)
    "retryOnTimeout": true,    // Retry when the provider times out
    "retryOnRateLimit": true,  // Retry when hitting rate limits (respects Retry-After header)
    "retryOnServerError": true // Retry on 5xx server errors
  },

  // --- @-file limits ---
  "atFileMaxFiles": 5,
  "atFileMaxFileSize": 51200,
  "atFileMaxTotalSize": 204800
}

MCP server config

Each server under mcpServers supports:

FieldTypeDescription
commandstringExecutable (for stdio transport)
argsstring[]Arguments passed to the command
urlstringHTTP endpoint (for streamable-http transport)
envobjectExtra environment variables for the server process
headersobjectCustom HTTP headers
transportstring"stdio" | "streamable-http" | "sse" (auto-detected if omitted)
preferredProtocolVersionstring"2025-11-25" | "2025-03-26" | "2024-11-05"
requestTimeoutMsnumberPer-request timeout
connectTimeoutMsnumberConnection timeout

Tip: Transport is auto-detected — if url is set without command, streamable-http is used. Otherwise stdio.

Environment variables

All settings can also be set via environment variables for CI / containers:

VariableSetting
DEEPSEEK_API_KEYAPI key (provider-specific vars also supported: OPENAI_API_KEY, ANTHROPIC_API_KEY, etc.)
AGENT_PROVIDERProvider override
AGENT_MODELModel override
AGENT_THINKING_LEVELThinking level override
AGENT_VISION_PROVIDERVision model provider
AGENT_VISION_MODELVision model ID
DSCODE_MAX_TOKENSMax tokens
DSCODE_CONFIG_HOMECustom config directory (default: ~/.dscode)
DSCODE_DATA_HOMECustom data directory
DSCODE_PROJECT_PATHProject directory
DSCODE_RETRY_MAX_RETRIESRetry max retries
DSCODE_RETRY_BASE_DELAY_MSRetry base delay
DSCODE_RETRY_MAX_DELAY_MSRetry max delay

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

DocumentWhat's inside
ARCHITECTURE.mdFull architecture: Agent as OS, 6-layer design, Driver/Skill model, source tree
ROADMAP.mdWhat's next: Sub-Agent system, System Prompt modularization, Diff-based editing
CONTRIBUTING.mdHow to contribute: philosophy alignment, OpenSpec SDD workflow, coding conventions
STYLE.mdTypeScript 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


Ver o original no GitHub

Bom para

Relacionados