DCS / Build & Agents / Agent Studio
Visual agent builder · receipts by default

Compose autonomous agents on a canvas. Every action signed.

Drag-and-drop multi-modal agents — text, vision, voice, action — from 641 base specialists. Wire them with conditional flows. Hit deploy. Every step writes a receipt to the same chain Anthropic and OpenAI don't own.

agent-studio · refund-agent.dcs FILES prompt.md agent.yaml tools.json tests/ > build me a refund-handling agent for Stripe PROMPT agent · refund-handler OUTPUT · agent.yaml name: refund-handler model: claude-sonnet-4 tools: [stripe_api, slack] guardrails: approval_above_500
Four agent kinds, one canvas

What you can build, end to end.

Every agent in Studio is built from one of four primitives. Compose them in series, in parallel, or in conditional branches. Multi-modal by design — pass an image into a vision agent, get text out, pipe it into a voice agent.

📝

Text agents

LLM-powered reasoning, drafting, classification, extraction. Backed by DCS Inference or any OpenAI-compatible endpoint.

  • Prompt + tools + memory
  • Stream or batch output
  • Per-message receipt
👁

Vision agents

Image and video understanding — OCR, captioning, object detection, layout extraction. Use stock vision models or bring your own.

  • Image, video, multi-page PDF
  • Bounding-box outputs
  • SBT-signed provenance
🎤

Voice agents

Real-time speech-to-text + text-to-speech for phone IVR, voice notes, transcription, agent-to-customer calls.

  • Twilio + WhatsApp ready
  • 11 languages out of box
  • Receipt per turn
🤖

Action agents

Tool-using agents that hit APIs, run scripts, write files, or call other agents. The autonomous layer.

  • 10+ built-in connectors
  • Custom tool spec
  • Spend cap + audit log
Visual canvas

Drag, drop, draw. No prompt-glue spaghetti.

Studio replaces "stack of prompts in a Notion doc" with a real graph: nodes are agents, edges are receipts, branches are if/else conditions. Test in the preview pane on the right.

studio.dcsai.ai/workspace/restaurant-pilot
● running
TRIGGER WhatsApp message 👁 VISION Read menu photo → items[], prices[] 📝 TEXT Parse intent → "order_pickup" DECIDE intent == ? 🤖 ACTION Place order in POS POST /api/orders 🎤 VOICE Callback & confirm tts(es-MX) SIGNED Receipt #4182 order question

Every edge in this graph emits a hash-chained, ed25519-signed receipt. The whole run is independently verifiable on the public verifier — without trusting your servers.

641 specialists

Compose with agents, not from scratch.

The agent library ships with 641 specialists across 28 industry verticals — restaurant, salon, clinic, real estate, logistics, retail, and more. Each is a black-box you wire into your canvas. Override any of them with your own prompts, tools, or models.

48
Page agents
590
Industry agents
37
Pipeline agents
9
Shared agents
1
Orchestrator

Browse the full 641-agent catalog →

Test → deploy → monitor

One click to a live endpoint. Two clicks to roll back.

Studio ships your composed agent as both an MCP server (Claude Desktop / Cursor / Windsurf) and a REST endpoint. Versioned. Receipt-signed. Reversible.

Run preview

Test the whole graph in the right-pane sandbox. Mock inputs, real outputs, no spend.

Deploy live

One click ships to an MCP server + REST endpoint. Versioned. SemVer tags. Side-by-side preview vs prod.

Roll back

Every deploy is a versioned receipt. Click any prior version in the timeline to revert in <30 seconds.

Pricing

Free to build. Pay only when your agent runs.

No seat fees. No usage caps on canvas time. You only pay for the model + compute your agents actually consume when deployed.

Hobby

$0/mo forever

For builders kicking the tires. One workspace, ten deployed agents.

  • Unlimited canvas time
  • 10 deployed agents
  • 10K agent runs / mo
  • Community receipts
Start free

Sovereign

Talkto us

Self-hosted Studio inside your perimeter. Receipts pin to your Filecoin allotment, not ours.

  • Air-gapped install
  • Your signing keys
  • SOC 2 + ISO 42001 ready
  • Annual contract
Sovereign details →
FAQ

Questions, answered straight.

Is this a no-code tool or a coder's IDE?
Both. The canvas is no-code by default — drag nodes, connect them, ship. For each node you can drop into a code panel and override the prompt, tool spec, or model. Power users live in the code; operators live on the canvas. Same graph.
What's the difference between Agent Studio and Agents API?
Studio is the visual builder. Agents API is the programmatic interface that the same composed agents expose once deployed. Build in Studio, call from your app via the API. See agents-api.html.
Can I bring my own model?
Yes — for any text or vision node, point at any OpenAI-compatible endpoint. Studio handles auth + retries + cost capping. If you're running on DCS Inference (Tier 0 H200), it's free for the first 10K runs / month.
How does the receipt chain work?
Every node execution produces an ed25519-signed receipt with the input hash, output hash, model used, cost, and a CID pointer to the prior receipt. Breaking the chain breaks the proof. The chain pins to Filecoin via Lighthouse so receipts outlive our servers. Read the full spec on DCS Provenance Standard v0.1.
Can I use Studio with Claude Desktop / Cursor / Windsurf?
Yes — the moment you click Deploy, the agent becomes an MCP server entry. Add the JSON snippet to your Claude/Cursor/Windsurf config and the agent appears as a tool inside that client. See MCP Server docs.
What happens if a node fails mid-flow?
Studio writes a failure receipt and either (a) rolls back to the last known-good state if you set the flow to atomic, or (b) continues down the error branch you defined. Both behaviors are part of the signed audit trail.

Stop gluing prompts. Ship a real agent.

Open Studio, drop a few nodes, hit Deploy. First agent live in under five minutes.

Start building free → Talk to a builder