Forma
ONNX graph editor for ML engineers
WHO USES ITML engineer · Developer converting models · Deep-learning learner
Use in the browser
Drop an .onnx file in the browser, see the graph, edit a node and export — no Python.
3 stars · checked on GitHub GitHub check 30 h late · no 7-day star increase observed (GH Archive) (as of ) · 2 Show HN points observed
NEW 1h OLD WHEN SEEN



TESIGN TAKE
The point is that graph editing and validation finish in the browser without your model ever being uploaded.
AT A GLANCE
- USAGE
- Unconfirmed — check before use.
- LANGUAGE
- TypeScript
- PLATFORM
- web
- ACTIVITY
- last commit 14 days ago () · latest release v2.5.0 () · 28 releases
- COMMUNITY
- 1 contributor · 0 open issues (incl. PRs) · made by: an individual
- SOURCES
- Show HN
- FIRST SEEN
- CATEGORY
- DEV TOOLS · AI
A summary, not legal advice.
WHY IT MATTERS
Most ONNX tooling stops at looking: you can read the graph, but changing one attribute sends you back to Python scripts. Forma renders the full computation graph when you drop an .onnx or .tflite file on the canvas, lets you click a node to edit attributes or rewire connections, and exports the result as a valid ONNX file. Everything runs in the browser via WebAssembly, so the model never leaves your machine; only the edit sequence is shared, as a link carrying the original file's SHA-256 fingerprint.
BUILD FROM THIS
- Pre-deployment surgery: turn a fixed batch size into a symbolic `batch` dimension or rename tensors
- A validation run that feeds the original and edited model identical inputs and compares outputs
- A minimal-repro ONNX file cut from just the nodes that misbehave (Extract Repro)
- A side-by-side text report comparing two model files' structure, attributes and latency
WHO IT'S FOR
- ML engineer
- fixes ONNX inputs, outputs and attributes right before deployment, without Python
- Developer converting models
- compares two files' structure before and after conversion
- Deep-learning learner
- clicks through a public model's layers and parameter counts
START IN 5 MINUTES
# 1. Open forma-ml.vercel.app
# 2. Drag any .onnx model file onto the canvas
# 3. Click any node to inspect it in the Layer Inspector panel
# 4. Make an edit and choose Share Edits to copy a verified edit link
# Run locally (Node.js 18+):
$ git clone https://github.com/Hussain004/Forma.git
$ cd Forma
$ npm install
$ npm run dev
# → open http://localhost:5173CAVEATS
- License unconfirmed — the README badge and the Show HN post say MIT, but the repository has no LICENSE file and GitHub metadata shows none
- ONNX and TFLite only; TFLite is read-only (no attribute editing, benchmark or Export Modified); convert .pt or .safetensors to ONNX first
- Desktop-first — a gate appears below 900px screen width
- Share links never contain the model, so the recipient needs the exact original file; INT8 sizes are analytical estimates
- No pricing stated; the web app opens without payment and the README links a donation page · shown on HN 2026-09-13 · the README lists releases up to v2.5.0 and 406 tests
RECEIPT
- FIRST SEEN
- AT SOURCE
- KEPT
- CREATED → FIRST SEEN
- 1h
- WHEN FIRST SEEN
- 3 ★
- SOURCES
- Show HN
The same facts in machine-readable form — View as Markdown · JSON
SIMILAR TOOLS
Up to five from DEV TOOLS by ★ total: edited entries first, then repository cards; this entry is highlighted. Drawn from the same stored snapshot as the rankings — a comparison, not a recommendation.
| IMAGE | NAME | ★ TOTAL | ▲ 7d | LICENSE | PLATFORM | LAST PUSH |
|---|---|---|---|---|---|---|
![]() | eccSkills, memory and security for agents | ★ 259k | ▲ +832 | MITpermissive | windows · macos · linux · cli | |
![]() | DeepSeek HarnessPlugin agent harness for developers | ★ 225k | ▲ +1,356 | MITpermissive | web · cli | |
![]() | ponytailMake your agent write less code | ★ 139k | ▲ +1,068 | MITpermissive | — | |
![]() | Spec KitSpec-first process for coding agents | ★ 137k | ▲ +560 | MITpermissive | cli · windows · macos · linux | |
![]() | FormaONNX graph editor for ML engineersTHIS ENTRY | ★ 3 | — | Licence unconfirmedunconfirmed | web |
TIMELINE
- Repository created
- Show HN post ↗
- FIRST SEEN BY TESIGN
- Published on TESIGN
SAME DAY
- 01
Hindsight
Narrative map for YouTube viewersRemembers 49,411 videos, maps 22,904 by issue, traces how a YouTuber's argument changed.
Use in the browser★ 0▲ 1 HN - 03
Words of Hanoi
Daily word puzzle for puzzle fansA daily puzzle: move letter tiles between five tubes until every full tube spells a word.
Use in the browser▲ 1 HN - 04
Qwen2.5 Program Slicing Security Auditor
JavaScript security audit for web devsPaste JS or HTML; a fine-tuned 7B model traces each risky sink to an exploit and verdict.
Use in the browser♥ 1 - 05
SubSweep
Subscription audit for AustraliansFinds forgotten subscriptions, price hikes and refund windows in an Australian bank CSV.
Use in the browser▲ 1 HN - 06
Spec Kit
Spec-first process for coding agentsMakes your coding agent write the spec first, then plan, build and verify against it.
Install to use★ 137k▲ +560 / 7d








