Hardware and Deployment Axis

Companion note · Addresses GitHub issue #8

---

Problem

clawfit's README promises an agent + LLM + hardware recommendation engine. The hardware field in the registry currently carries three values (laptop, workstation, cloud) which are functional but coarse. Several important distinctions are collapsed:

• A developer's M3 MacBook and a $3,000 Linux workstation with an RTX 4090 are both hardware: workstation in different contexts
• cloud conflates managed APIs (Anthropic/OpenAI), self-hosted cloud VMs, and vendor-managed platforms (AWS Bedrock, Azure AI, Vertex AI)
• There is no axis for governance requirements (data residency, VPC, HIPAA, SOC 2)
• There is no axis for isolation model (process-level, container, microVM, k8s)

---

Proposed hardware/deployment taxonomy

Primary hardware categories (current + expanded)

Category	Description	Inference substrate
laptop	Consumer notebook; ≤64 GB RAM; Apple Silicon or mid-range x86	Ollama, llama.cpp, MLX
workstation	Desktop or tower; dedicated CUDA/AMD GPU; 24–80 GB VRAM	Ollama, vLLM (local), llama.cpp
edge	Mobile, embedded, IoT; ARM; ≤8 GB RAM	LiteRT-LM, MLC LLM (iOS/Android)
cloud_api	Managed API endpoint; no user-controlled infra	Anthropic, OpenAI, Mistral, DeepSeek
cloud_vm	Self-hosted on rented VM/GPU; user manages runtime	vLLM, TGI, TensorRT-LLM on EC2/GCP
cloud_managed	Fully managed platform with governance surface	AWS Bedrock, Azure AI, Vertex AI
home_cluster	Multiple consumer devices networked together	exo

Deployment context dimensions

These are independent of hardware and should be filterable:

Dimension	Values	Relevance
governance_need	none / basic / hard	HIPAA, SOC 2, data residency requirements
isolation_model	process / container / microvm / k8s	Security boundary required
managed_platform	bedrock / azure_ai / vertex_ai / none	Which managed surface hosts the model
multi_tenancy	single / team / enterprise	Concurrent user isolation requirements

---

How this maps to existing clawfit filters

The current hardware filter collapses the above into three choices. Proposed evolution:

Current:   hardware: [laptop | workstation | cloud]

Expanded:  hardware_type:  [laptop | workstation | edge | cloud_api | cloud_vm | cloud_managed | home_cluster]
           governance:     [none | basic | hard]
           isolation:      [process | container | microvm | k8s]

Backward compatibility: The three current values map directly:

• laptop → hardware_type: laptop
• workstation → hardware_type: workstation
• cloud → hardware_type: cloud_api (default assumption)

---

Connection to inference runtime substrate (#9)

The hardware axis and the inference runtime substrate axis (issue #9) are complementary:

Hardware context	Inference substrate default
laptop (Apple Silicon)	MLX or Ollama (Metal backend)
laptop (x86)	Ollama or llama.cpp (CPU/CUDA)
workstation (CUDA)	Ollama + GPU, vLLM, or llama.cpp
cloud_vm	vLLM, TGI, TensorRT-LLM
cloud_api	Provider-managed; substrate is opaque
cloud_managed	Bedrock/Azure/Vertex — substrate is opaque + auditable
home_cluster	exo (multi-device sharding)

---

Discriminating criteria for hardware classification

Use cloud_api when: the agent calls a remote API endpoint; no user-controlled inference; latency is network-bound; network: online is required.

Use cloud_managed when: governance, data residency, or compliance (HIPAA/SOC 2) drive the platform choice; the managed surface (Bedrock/Azure AI/Vertex AI) provides audit trails and VPC isolation.

Use cloud_vm when: the user controls the inference stack on rented compute; vLLM or TGI is deployed on EC2/GCP/Azure VMs; cost is usage-based but infra is self-managed.

Use workstation when: dedicated local GPU (NVIDIA/AMD) is available; VRAM ≥ 16 GB; user can run 7B–70B models without quantization.

Use laptop when: consumer notebook; VRAM < 16 GB or no discrete GPU; quantized models (GGUF Q4/Q5) are the realistic inference path.

Use edge when: mobile or embedded target; ARM processor; battery / thermal constraints limit sustained inference.

Use home_cluster when: user has multiple consumer devices and wants to shard a model too large for any single device.

---

Recommended v0.4 schema change

Add to clawfit/registry/hardware.json entries:

{
  "id": "cloud-managed",
  "name": "Cloud Managed Platform",
  "type": "cloud_managed",
  "governance_profiles": ["hipaa", "soc2", "gdpr"],
  "managed_platforms": ["bedrock", "azure_ai", "vertex_ai"],
  "isolation_model": "k8s",
  "network": "online"
}

Current hardware.json has 5 entries. Suggested expansion to 7 (add cloud_managed and edge).