TongFlow Is Now a DeepSeek Harness Plugin — Plus Everything Since v0.2.1

The biggest change since our last roundup is not a model — it’s where TongFlow runs. TongFlow is now installable as a plugin for DeepSeek Harness (dsh), DeepSeek’s open-source agent harness. Below that headline sits a month of smaller releases (v0.2.2 → v0.3.3): a form-style App Mode, undo/redo, connections you draw by hand, one-key router plugins with a per-node model picker, and eight new official plugins.

New to TongFlow? It’s an open-source multi-modal GenAI workflow studio — every model is a node on an infinite canvas. Start with the introduction.

dsh-tongflow: a film crew inside your agent

Install it with one line:

npx @deepseek-ai/dsh@next plugin --profile web add dsh-tongflow

Then start a chat whose first message begins with @tongflow. That session turns into a Studio: the conversation on the left, your project tree in the middle, the preview or the TongFlow canvas on the right — no separate TongFlow server, the canvas is embedded in dsh’s own web UI.

The division of labour is strict, and that is the point:

• dsh is the harness — sessions, models, tools, background jobs, the web shell.
• The agent does the creative work — treatment, script, characters, shot list, review notes. Plain files, in your language.
• TongFlow does the deterministic work — every image, voice line, music cue and video clip is produced by running a workflow file (*.tongflow.json) through TongFlow plugins.

There is deliberately no “generate an image” tool. The agent writes one workflow per asset, saved right next to the takes it produces, binds its inputs to project assets (tf://CHR_MEI/REF, tf://EP01_SC003_SH0010/dialogue/2), runs it, looks at the result, and circles the good take. You can open the same file on the canvas, tweak a node, and re-run. When a shot is done, one click composes its asset workflows into a single big workflow — so a human can confirm and fine-tune the whole chain before the episode renders.

The project on disk is organised the way a real crew works: story/, world/ (characters, locations, props, a style bible with a consistency kit), episodes/, shots/, notes/, export/. Shots are EP01_SC003_SH0010, takes are T01…, and every take carries a provenance.json — which workflow, which bindings, which plugins, how long. A manga-drama template ships first (script → character sheets → storyboard → keyframes → dubbing → image-to-video → assembly), with English and Chinese starter files.

Under the hood this is the same TongFlow engine you use in the studio, driven through the new tongflow npm package (the framework-free workflow core plus a tongflow/canvas React entry) and the Python SDK 0.3.0. Any project that wants an agent to build and run workflows can use the same two pieces.

Where to find it: packages/dsh-tongflow in the TongFlow repo, dsh-tongflow on npm, and soon on the awesome-dsh-plugin list.

One honest note: v0.3.0 shipped a chat panel inside the TongFlow canvas that built workflows for you. It is gone in v0.3.3. Agent-driven building belongs in an agent harness, and dsh-tongflow does it better than a side panel ever could — TongFlow itself stays a pure workflow product.

The canvas got friendlier

• App Mode (v0.3.0) — present any workflow as a simple form: inputs on top, results below, canvas hidden. A finished workflow doubles as a reusable tool.
• Undo / redo (v0.3.0) — Cmd+Z / Cmd+Shift+Z, with typing coalesced into single steps.
• Manual connections (v0.3.2) — drag an edge between any two handles; the contract is checked live while you drag, so only valid connections land.
• One-click auto-layout and consistent node spacing (v0.3.1).
• Settings overhaul and first-run onboarding (v0.3.0) — readable cards, a guided banner, one-paste Modal token setup, and failures that come with an error code, a localized message and a link to the field you need to fill.
• One-click runs handle dynamic splits (v0.2.3) — split a document into N parts and every downstream node runs N times, in order. Previously only the first item was processed.
• Open-vocabulary sound separation (v0.2.2) — a separate-sound node: describe a sound in words (“dog barking”) and split it from everything else, served by SAM-Audio or fal.
• Omni-reference video (v0.3.1) — select any mix of images, clips and audio and the smart island offers refs-gen-video; address them as <Picture 1> / <Video 1> / <Audio 1> in the prompt.

Router plugins: one key, dozens of models

The aggregator plugins were split out as tongflow-router-* and each got a per-node model picker: Replicate, fal.ai, APIMart and OpenRouter. Paste one API key and choose FLUX, Seedream, Veo, Kling, Whisper or Hunyuan3D right on the node.

Eight more official plugins

• Krea 2 Turbo — open-weights 12B text-to-image, 8 steps, up to 2K.
• Qwen3.8-27B — multimodal text: image and video understanding, transcription, on your own Modal GPU.
• IndexTTS-2.5 — zero-shot voice cloning with emotion control (“emotive speech”).
• MiniMax-Music3 — self-hosted music generation.
• SenseNova-Vision — visual QA, detection and OCR, surface normals, matting, pose.
• DeepSeek — V4 flash / pro with thinking on or off, and a live thinking bubble beside the node while it reasons.
• xAI and Runway — API plugins.
• And, covered in its own post: MiniMax-H3 video with native stereo audio.

Also new: the tongflow-local-* prefix for plugins that run on your own machine — the first one runs MiniMax-H3 natively on Apple silicon.

The catalog now stands at 45 official plugins. LTX and FastWan were retired; the built-in example workflow uses MiniMax-H3 for its image-to-video step.

Try it

• Studio: app.tongflow.com, or grab the v0.3.3 desktop shell from Releases.
• Inside your agent: npx @deepseek-ai/dsh@next plugin --profile web add dsh-tongflow, then say @tongflow and describe the show you want to make.
• Build on it: npm i tongflow · pip install tongflow.

If TongFlow is useful to you, a star on GitHub helps a lot.

Expand your imagination, stretch your ideas — give it a try.