EvoSprig 0.1.8 — Desktop Studio

Desktop Studio is available for Windows, macOS and Linux, with the Python backend included.

• Build with the visual canvas or Python, debug nodes, and export independent applications.
• Connect local models, your own inference API or a cloud provider. Define model roles during development and bind concrete services at deployment, with shared defaults and per-node overrides.
• Preload local services on selected GPUs and reuse them across tasks.
• Use adjustable workspace panels, customer-managed knowledge libraries, shared memory and scoped self-evolution.
• Explore Episode Studio, comics, office workflows and three hands-on tutorials in English and Chinese.

Downloads: Windows .exe; macOS .dmg (Apple silicon / Intel); Linux .AppImage or .deb. Python distributions and checksums are included below.

Installation guide · 中文安装指南 · Model deployment · Tutorials

v0.1.7 - Agent harness evolution

Evolve the parts that shape how an agent works: prompts, skills, code, workflow graphs, model choices and per-agent harness settings.

• Nine new native methods: SkillOpt, Evo-Harness, AHE, Meta-Harness, RRSI, ADAS, DarwinX, ModularRSI and Arbor — 23 strategies across 21 families.
• Developer-defined search spaces, complete candidate execution, resource/graph diffs and standalone exports.
• Persistent peer messages, replies, acknowledgements and resumable waits for Cowork teams.
• Studio controls for agent harnesses, memory retrieval and structured search.

Start with the harness search guide, CoworkOS (EvoSprig) or the optimizer plugin guide.

v0.1.6 — Cowork workbench and industry environments

A clearer view of what you can build with EvoSprig.

• Four featured applications: VideoClaw, CoworkOS, RAI robot manipulation and industry environment construction, with direct links to the original projects.
• Cowork workbench: conversations on the left, work in the center, and team members and outputs on the right. Includes light/dark themes, mobile panels and review-draft preservation.
• Typography and diagrams: bundled open-licensed fonts, refreshed canvas screenshots, and a detailed architecture and capability map.
• Industry development: a complete guide from business scenarios through environment construction, validation, task evaluation and agent self-evolution.

Applications · CoworkOS / CoworkRSI · Industry environments

v0.1.5 — Robot manipulation and CoworkRSI

Build visual robot agents and collaborative teams that learn from reviewed work.

• RAI manipulation (EvoSprig): a multimodal Franka tabletop agent, CPU simulation tools, recoverable action receipts and physical task checks. Watch sorting and stacking in the application gallery.
• CoworkOS (EvoSprig): the main agent checks completed work and sends targeted corrections back to specialists.
• CoworkRSI: retrieve lessons before planning and save a retrospective after human review; improve selected team instructions using your evaluation tasks.
• Presentation: compact VideoClaw film cards, clear application names, homepage News and an updated hero.

Robot application · CoworkOS / CoworkRSI · VideoClaw gallery

v0.1.4 — CoworkOS, video gallery and custom evolution methods

Build a team that works with you, then improve its instructions on your own tasks.

• CoworkOS, built with EvoSprig: a main agent coordinates specialists, produces shared documents, accepts edits and resumes scheduled follow-ups. Includes an English/Chinese workbench.
• VideoClaw gallery: Moon Garden, The Last Library and Cloud Bakery, with animated previews and complete films.
• Custom evolution methods: register an optimizer in Python or an installed package and reuse task execution, scoring, budgets and version selection.
• Runtime improvements: revision-aware write feedback, reported-usage settlement on truncated responses, isolated media model workers and explicit job retries.
• Documentation: separate guides for application developers and framework contributors, a component/method coverage map, and larger architecture graphics with an editable slide.

Applications · Build with canvas and Python · Add an algorithm

v0.1.3: Vision and local video production

Build video-producing agents with native image inputs, Kimi visual review, and local FLUX.2-klein / HunyuanVideo-1.5 tools.

• Generate images and videos through durable HTTP jobs, with idempotency, cancellation, retained assets and FFmpeg composition.
• Select image inputs on model nodes and preview images or videos in Studio's results and approval steps.
• Reuse character references, review individual frames, create replacement takes and export the workflow as a standalone application.
• Connect the official VideoClaw workflow to the same language, image and video models through the included adapter.

Watch the two short films and the frame-retake comparison in the video studio example. See local media setup and the vision guide.

Model weights are downloaded separately under their respective licenses.

EvoSprig 0.1.2 — Self-evolving agents, clearer workflows

EvoSprig 0.1.2 makes self-evolution easier to inspect: application pages show real prompt and skill edits beside their outputs, and Studio uses distinct node cards with clearer connections and nested groups.

• Add MIPRO resource search: grounded instruction proposals, successful component demonstrations and categorical TPE. The registry now offers 14 training-free strategies across 12 families.
• Reuse model HTTP connections across nodes and evaluation batches, with configurable connection limits.
• Add bounded judge-output repair and include actual questions and human replies when evaluating interactive applications.
• Refresh 20 workflow canvases and 14 application result images, including teaching and book-writing comparisons.
• Simplify the architecture diagram and provide its editable PowerPoint version.

Start with the application gallery, evolution guide or quickstart.

The attached wheel, source distribution and source ZIP include the framework and examples. Dependency inventory and SHA-256 checksums accompany the release.

EvoSprig v0.1.1 — Prompts, skills and application workflows

Build an agent, attach reusable skills, add task inputs and choose an evolution method. Studio now keeps prompt and skill editing, tool access, memory and model parameters together in the node inspector, and shows resource changes beside application outputs.

• Ordered skill resources with independent optimization; unchanged workflows keep their canonical hashes.
• Per-node temperature, top-p, output limits and tool controls.
• Career coaching, multi-channel content and interactive learning applications built on EvoSprig.
• Shorter application pages with results shown inline, refreshed canvases and an editable architecture slide.

See the agent configuration guide and application gallery.

EvoSprig 0.1.0

EvoSprig's first open-source release brings visual workflows, persistent memory, application evaluation and controlled evolution into one Python platform.

• Build on a clearer canvas: distinct node shapes and capability icons, layered layout, nested groups, collapse/expand, and executable subflow editing. Example projects include portable Studio views.
• Develop real applications: 17 illustrated guides covering software teams, data analysis, research, receipts, video production, writing, planning, support, memory and industry environments.
• Improve with your own tasks: 13 training-free strategies across 11 method families, input-only LLM judging, custom checks, bounded budgets and reviewed releases. Named methods include original paper references.
• Evaluate industry workflows: connect an IndustryBench installation for isolated business sessions, tool discovery and business-state replay scoring. A native seven-stage workflow produces environment drafts and pauses for review; integrate and validate the drafts before using a new environment.
• Tune execution: per-node output limits, bounded format repair and configurable concurrent evaluation with isolated sample identities.
• Take the application with you: standalone Python, HTTP and container build projects; English/Chinese Studio and READMEs.

Start with the quickstart, browse the applications, or read the industry guide.

The source archive includes Studio source. Build it with Node.js 20+; the Python wheel requires Python 3.12+. Optional model, server and optimizer dependencies are installed through the documented extras. IndustryBench and media providers are separately configured integrations.