AI 에이전트 + LLM + 하드웨어 추천 엔진 — 58개 도구, 7레이어 에코시스템 맵, 10차원 조직 적합도 스코어링

Agent + LLM + hardware recommendation engine and evidence hub.

Read in: 한국어 🇰🇷

clawfit answers a practical question:

Given a task, latency target, budget, network conditions, and team maturity, what combination of agent pattern, model, and hardware is the best fit?

It is three things in one:

1. Recommendation engine — 58 tools scored across 10 dimensions (task fit, maturity, role, layer relevance, team size, network, latency, features, complexity, budget). Hard multipliers penalize critical mismatches (offline required + online-only tool → x0.25).

2. Ecosystem map — A 7-layer taxonomy (L1 base runtimes → L7 human interfaces) with 70+ research-watch signal documents tracking GitHub Trending, GeekNews, and Hacker News daily via automated agents.

3. Org-fit diagnosis — A 10-question interactive questionnaire (TUI, CLI, or web) that builds an organization profile and recommends a prioritized multi-layer tool stack.

• Teams choosing between agent stacks (Claude Code vs OpenClaw vs Aider vs ...)
• DevOps deciding local vs cloud execution topology
• Executives evaluating AI tool adoption strategy
• Researchers mapping the AI agent ecosystem
• Anyone building an evidence-backed recommendation layer

Full analysis in docs/research-watch/ (70+ docs) · Full map in docs/reference-levels.md

Option A — pipx (recommended, globally available, no venv needed)

pipx install git+https://github.com/hongsw/clawfit

Option B — editable install (for development / hacking)

git clone https://github.com/hongsw/clawfit.git
cd clawfit
python3 -m venv .venv && source .venv/bin/activate
pip install -e .

Answer 10 questions about your team → get a prioritized multi-layer recommendation.

TUI (recommended — navigate with arrow keys, results update live in split pane):

clawfit tui

 ████████████░░░░░░  5/10  [USECASE]
 ──────────────────────────┬──────────────────────────────
 What is the main thing    │ Stage 4 — Tool-using agent
 you want AI to do?        │
                           │ [PRI] L1 Base runtime
  ○ Write or review code   │    45% Claude Code
  ● Research & summarize   │    39% Aider
  ○ Answer questions (QA)  │    38% Goose
  ○ Classify / route data  │
  ○ Analyze data           │ [PRI] L4c Tool-use infra
  ○ Summarize at scale     │    41% Serena
                           │    35% Context7
 ─ answered ─              │
  Team size: small team    │ NEXT STEP
  Role: developer          │ You're ready for a meta-wrapper...
 ──────────────────────────┴──────────────────────────────
  ↑/↓ Move   Space/Enter Select+Next   ← Back   → Next   q Quit

CLI (non-interactive, pass answers as JSON):

clawfit diagnose --answers '{
  "team_size": "small",
  "primary_role": "developer",
  "current_ai_usage": "coding_agent",
  "primary_task": "code-gen",
  "output_destination": "team",
  "frequency": "daily",
  "data_sensitivity": "internal",
  "monthly_budget": "medium",
  "governance_need": "soft",
  "growth_horizon": "deepen"
}'

Web UI (browser with live filtering):

clawfit serve          # opens http://localhost:7771
clawfit serve --port 8080

clawfit recommend --task qa --latency low --budget 0.01

clawfit recommend \
  --task code-gen \
  --latency medium \
  --budget 0.01 \
  --hardware cloud \
  --network online \
  --statefulness session \
  --maturity 5 \
  --top 5

--maturity 5 = sub-agent user stage. See the maturity × layer map for all 11 stages.

clawfit list agents
clawfit list llms
clawfit list hardware
clawfit profile

10-dimension weighted scoring with hard multipliers:

Hard multipliers (applied after weighted sum):

• Offline required + online-only tool → x0.25
• Role mismatch (no role overlap) → x0.75

The pipeline is intentionally simple and inspectable:

1. Registry loading — load 58 tool definitions with 10-field org_fit metadata
2. Profile building — convert 10 questionnaire answers into an OrgProfile
3. Scoring — score each tool across 10 dimensions + hard multipliers
4. Layer grouping — group by ecosystem layer (L1–L7), prioritize by maturity stage
5. Recommendation output — return prioritized multi-layer stack with rationale

clawfit/
├─ .claude/agents/          ← harness team sub-agents (5)
├─ clawfit/
│  ├─ cli.py                ← argparse CLI (recommend, list, tui, serve, diagnose)
│  ├─ org_scorer.py         ← 10-dimension scoring engine
│  ├─ tui.py                ← curses TUI with split-pane live preview
│  ├─ server.py             ← stdlib HTTP server (localhost:7771)
│  ├─ diagnose.py           ← interactive CLI questionnaire
│  ├─ filters.py            ← hard constraint elimination
│  ├─ scoring.py            ← cartesian product scoring (agent × LLM × hardware)
│  ├─ recommend.py          ← public API: recommend() → list[dict]
│  ├─ schemas.py            ← dataclasses: Agent, LLM, Hardware, Recommendation
│  ├─ loader.py             ← loads registry/*.json
│  ├─ data/
│  │  ├─ tools_registry.json  ← 58 ecosystem tools with org_fit (10 fields each)
│  │  └─ org_questions.json   ← 10-question bank, 3 phases
│  └─ registry/             ← agents.json, llms.json, hardware.json
├─ docs/
│  ├─ reference-levels.md   ← ecosystem map v0.3 (7-layer taxonomy)
│  ├─ research-watch/       ← 70+ signal analysis docs (daily scan)
│  └─ pages/                ← ecosystem-overview, ecosystem-axes, maturity-layer-map
├─ data/
│  └─ tools_registry.json   ← mirror of clawfit/data/
├─ tests/
│  ├─ test_filters.py
│  └─ test_recommend.py
└─ pyproject.toml

clawfit tracks a broader AI tooling landscape documented in:

• docs/reference-levels.md — canonical 7-layer ecosystem map
• docs/pages/ecosystem-axes.md — classification logic, boundary rules, worked examples
• docs/research-watch/ — 70+ individual tool/trend analysis documents (daily automated scan)
• docs/pages/maturity-layer-map.md — how user maturity stages (1–11) map to tool layers (L1–L7)

from clawfit.recommend import recommend

results = recommend(
    task="research",
    latency="high",
    network="online",
    top_n=3,
)

print(results[0])

python -m pytest tests/ -v

Contributions are welcome, especially around:

• registry expansion (new tools with complete org_fit metadata)
• scoring logic improvements
• benchmark references and evidence
• research-watch signal analysis

Open an issue or PR with: what you are adding, what evidence supports it, and how it fits into the comparison model.

MIT