CLI Reference

The manta binary provides core workflow commands plus several operational helpers.

manta --help

Commands

Command	Purpose
info	Show model info: tensor names, dtypes, shapes, and capacity
capacity	Measure embedding capacity per tensor (JSON output)
embed	Embed a payload into a model (safetensors or GGUF)
extract	Extract a payload from a model
metrics	Compare original vs modified tensors and rank perturbation impact
campaign	Generate a campaign config from a deployment profile
model-card	Generate a HuggingFace model card for cover
payload	Build typed payloads (c2, beacon, creds, exfil, raw)
key	Encode, decode, or inspect extraction keys
deliver	Package and upload weaponized models to HuggingFace Hub with auto_map hook generation, or to generic HTTP registries
split-payload	Shamir-split a payload into N shares (K-of-N threshold)
combine-payload	Reconstruct a payload from K Shamir shares
B --> C[embed]
C --> D[metrics]
D --> E[extract]
C --> F[campaign]
C --> G[model-card]
C --> H[deliver]
B --> I[split-payload]
I --> C
E --> J[combine-payload]
```

1. Recon — info and capacity to assess the target model
2. Weaponize — embed to insert payload, metrics to quantify perturbation
3. Deliver — campaign to generate deployment config, model-card for cover, deliver to upload (with optional trust_remote_code hook for HuggingFace)
4. Exploit — extract on the target to recover payload
5. Multi-model — split-payload before embedding, combine-payload after extracting from K models

Global Options

Flag	Description
--version	Print version
--help	Print help