Toyhom/EvoSprig

News

• 2026-11-06 · Desktop Studio and a broader development workflow — Windows, macOS and Linux clients; scoped self-evolution; updatable RAG libraries and shared memory; three hands-on tutorials.
• 2026-11-04 · ComfyUI workflows — import workflow JSON, edit and debug its tool nodes on the canvas, and export back to ComfyUI.
• 2026-11-04 · Episode Studio — a multi-agent production team that turns stories into animated films, with a recurring cast and automatic shot review.
• 2026-11-02 · EvoSprig is open source: a platform for building and evolving agent applications with Python or a visual canvas.

Build agents with Python or a visual canvas, then improve their behavior automatically on your own tasks. EvoSprig combines multi-agent workflows, tools, persistent memory and controlled self-improvement (RSI): generate candidate versions, evaluate the running application, compare results and enable a version. Export your application to run independently.

English narration · 中文配音

What will you build?

Episode Studio — a story, a cast, a film

A production team turns a novel or one sentence into an animated film. Agents choose the style, write the script, create the cast and voices, and direct and review each shot.

VideoClaw (EvoSprig) A story idea becomes a short film. Original: VideoClaw	CoworkOS (EvoSprig) A team that plans, writes and works with you. Original: CoworkOS
RAI manipulation (EvoSprig) See the scene. Move the objects. Check the result. Original: RAI	Industry environment builder Build a business world for your agent to learn in. Environment guide
Comic Studio A story becomes a comic, with a shared cast and panel-by-panel review. Build it in ten minutes	Office deliverables One table becomes an editable spreadsheet, report and slide deck.
Circuit Lab Describe a circuit. Get a netlist, a simulation and measured results.	Trade follow-up Turn customer conversations and product facts into a useful reply.

Build an agent that…	Start here
Directs an animated story with dialogue	Episode Studio — a shared cast, connected shots and automatic retakes
Repairs code and returns a tested patch	Super-agent (EvoSprig) — a repository team based on DeepSeek's Super-agent
Turns a story into a film	VideoClaw (EvoSprig) — characters, scenes and a finished cut (original VideoClaw)
Works with you on a project	CoworkOS (EvoSprig) — specialist teams and shared documents (original CoworkOS)
Sees a scene and controls a robot arm	RAI manipulation (EvoSprig) — visual planning, pick-and-place and task checks (original RAI)
Builds a business sandbox for agent development	Industry environment builder — procedures, tools, data and business-state checks
Tailors a resume and runs interview practice	Career coach — job fit, application materials and feedback
Writes a Xiaohongshu post from your notes	Xiaohongshu Studio — reader angles, style examples and guided revisions
Produces editable office files	Office deliverables — spreadsheet, brief and slides from one table
Designs and tests a small circuit	Circuit Lab — netlists, ngspice waveforms and measured targets
Drafts a useful buyer follow-up	Trade follow-up — customer history, product facts and reviewed drafts
Draws a complete comic page	Comic Studio — shared cast, panel review and page layout
Turns one announcement into a content campaign	Content studio — parallel channels and shared editing skills
Teaches a topic and reviews your answers	Study coach — lessons, exercises and personal feedback
Turns a brief into a working web app	Software team — design, implementation, browser checks and repair
Runs Python and explains your data	Data Interpreter — files, calculations and charts
Helps you make a difficult decision	Decision room — opposing arguments, rebuttals and a moderated recommendation
Reads receipts and checks expenses	Receipt assistant — OCR, structured extraction and policy review
Researches a question with a team	Research team — parallel investigation, sources and editorial approval
Organizes the rest of your work	Meetings · Books · Travel

Also explore SQL analysis, document chat, customer support and personal memory. Open the gallery for setup, workflow canvases and application outputs.

Why EvoSprig?

• Self-evolving agents, under your control. 23 training-free strategies across 21 method families, including SkillOpt, ADAS, ModularRSI, DarwinX and Arbor. Evolve prompts, skills, harnesses, workflows and code. Choose the resources, budget and release criteria.
• Evaluation that fits your application. Start with task inputs and an LLM judge, or bring reference answers, business checks and a sandbox. Score outputs, steps, tools and environment outcomes.
• Code and canvas, one application. Build typed workflows with parallel agents, dynamic map, loops and human approvals. Organize large teams with nested groups and edit executable subflows.
• Build the environment as well as the agent. Turn a finance, government or manufacturing scenario into schemas, procedures, tools and data drafts. Connect the validated business sandbox to task evaluation and self-evolution. Get started.
• Knowledge that stays current. Import PDFs, Word files and scans; combine keyword and vector retrieval. Let customers update the deployed knowledge library. Add scoped facts, shared context and persistent memory. Build a RAG application.
• Tools for real work. Vision inputs, MCP, HTTP, SQL and Python plugins; optional code execution, browser checks and OCR; read-only concurrency and recoverable tool feedback.
• Run, recover and ship. Checkpoints and external job receipts preserve progress. Export standalone Python, HTTP applications or container build projects.

How it fits together

Build the application on the canvas or in Python; both produce the same versioned AgentSpec. The runtime coordinates agents, tools and shared memory, saving progress for debugging and recovery. Evaluation feeds real tasks and outcomes into controlled evolution, which changes selected resources or workflow settings; compare the candidates, enable a version and export the application. ComfyUI graphs can join the workflow as editable tools and be exported back to ComfyUI.

Get started

Desktop: Windows, macOS and Linux · Web: run Studio in your browser.

Requires Python 3.12+. Studio also requires Node.js 20+.

git clone https://github.com/Toyhom/EvoSprig.git
cd EvoSprig
python -m pip install -e '.[server,evolution,mcp]'

# Run your first workflow without an API key.
evosprig run examples/hello/agent.yaml --input '{"question":"Hello"}' --mock

# Open Studio.
npm --prefix apps/studio ci
npm --prefix apps/studio run build
evosprig serve --project examples/hello/agent.yaml \
  --studio-dir apps/studio/dist --workspace .evosprig

Open http://127.0.0.1:18766. Studio starts in English; switch to 中文 in the sidebar. Your language preference is saved in the browser.

For live models, configure environment-based credentials in a deployment profile. Follow the quickstart for model connections and export.

Self-improvement you can see

Build your agent, add task inputs, select the resources to change and choose an algorithm. EvoSprig writes candidates and runs your application to compare them. Review the prompt or skill changes alongside the outputs, then enable your chosen version. Configure your first improvement.

In the study coach, one Self-Refine round changed two shared skills so exercises and feedback stay tied to the supplied lesson. The page shows the actual edits and answers, using tasks adapted from GSM8K.

Explore the book-writing comparison and document-answering comparison to see prompt changes and selected outputs. Each page shows what changed and how it affected the answer.

Bring your own inputs to LLM-as-judge, or connect your evaluator. The method guide explains which feedback each strategy uses.

Documentation

Follow the tutorials to build your first application. Use the reference to look up models, tools, memory, evaluation and deployment. Both are available in Chinese.

Build applications: canvas and Python · Extend the framework · Add an evolution algorithm

Workflows and architecture · Memory · Tools and MCP · Durable execution · Deployment · Roadmap

Contributing

Bug reports, example applications and integrations are welcome. Read the contributing guide to run checks and submit a change.

License

Apache-2.0.