Source Repo Analysis

Source Repo Analysis

Last Mile 360 was built by studying 16+ open-source projects in the AI agent and code analysis space. For each project, we evaluated what patterns and concepts were worth adopting — and explicitly rejected importing any code or dependencies. This page documents every evaluation.

Key principle: Patterns are portable. Dependencies are liabilities.

---

The Claw Family (Inference Engines)

OpenClaw

• Stars: 247k+ | Language: TypeScript | Risk Level: 🔴 CRITICAL
• What it is: The dominant open-source LLM inference engine
• What was taken: Nothing — no code, no dependency, no import
• What was rejected: The entire runtime. Self-hosted inference means managing GPU servers, CUDA drivers, model weights, and OOM kills
• Why rejected: Last Mile uses Claude API + Workers AI instead. Zero infrastructure to manage, predictable costs, automatic scaling. For a security product, running your own inference cluster is an unnecessary attack surface
• Replacement: Tier 1 Claude API (complex analysis) + Tier 2 Workers AI (edge inference)

NanoClaw

• Stars: 11k+ | Language: TypeScript | Risk Level: 🔴 CRITICAL
• What it is: Lightweight variant of OpenClaw for smaller models
• What was taken: Nothing
• What was rejected: Same reasoning as OpenClaw — self-hosted inference is operational overhead with no benefit for our use case
• Why rejected: Workers AI already provides lightweight models at the edge with zero cold start
• Replacement: Cloudflare Workers AI

PicoClaw

• Stars: 12k+ | Language: Go | Risk Level: 🟠 HIGH
• What it is: Minimal Go implementation for model serving
• What was taken: Concept only — the idea that inference can be decomposed into small, focused units
• What was rejected: All code. Go runtime adds another language to the stack
• Why rejected: Cloudflare's platform is JavaScript/TypeScript-native; adding Go increases complexity without adding capability

ZeroClaw

• Language: Rust | Risk Level: 🟠 HIGH
• What it is: Rust-based inference with WASM compilation target
• What was taken: Concept only — WASM as a portable execution target is interesting for future Workers integration
• What was rejected: All code. Rust WASM compilation toolchain is complex
• Why rejected: Premature optimization. Workers AI handles inference today; custom WASM inference is a Phase 6+ consideration

MimiClaw

• Language: Pure C | Target: ESP32 | Risk Level: 🔴 CRITICAL
• What it is: Runs tiny models on microcontrollers
• What was taken: Nothing — zero overlap with cloud-based code analysis
• What was rejected: Everything. Embedded C on ESP32 has no relevance to web security scanning
• Why rejected: Completely different domain. Listed for completeness as part of the Claw ecosystem evaluation

---

Agent OS / Multi-Agent Frameworks

OpenFang

• Stars: 14k+ | Language: Rust | Risk Level: 🟡 MEDIUM
• What it is: Agent operating system with strong isolation primitives
• What was taken: Agent isolation pattern — the concept that each agent should run in its own sandboxed environment with no shared memory or filesystem
• What was rejected: Rust codebase, custom runtime, process management
• Why relevant: Directly influenced the decision to run each Last Mile agent as a separate Cloudflare Worker (V8 isolate = natural sandbox)
• Implementation: Each agent is a separate Worker deployment with its own Queues binding

CrewAI

• Language: Python | Risk Level: 🟡 MEDIUM
• What it is: Framework for orchestrating role-based AI agent teams
• What was taken: Role-based crew pattern — the idea that agents should have explicit roles, goals, and backstories that constrain their behavior
• What was rejected: Python runtime, LangChain dependency, custom agent loop
• Why relevant: Influenced agent design: each Last Mile agent has a defined role (Security, Database, Infra, Observability, Quality) with specific rules it's responsible for
• Implementation: Agent specialization via rule ownership, not general-purpose reasoning

AutoGen (Microsoft)

• Language: Python | Risk Level: 🟢 LOW
• What it is: Multi-agent conversation framework
• What was taken: Conversational agent pattern — agents can discuss findings and cross-reference results
• What was rejected: Python runtime, complex conversation management, Azure dependencies
• Why relevant: Informed the results aggregation phase where findings from different agents are correlated (e.g., a missing RLS policy found by Database Agent affects the Security score)
• Implementation: Cross-agent correlation in the scoring aggregator

SuperAGI

• Language: Python | Risk Level: 🟠 HIGH
• What it is: Autonomous AI agent framework with tool use
• What was taken: Concept only — the autonomous execution loop is interesting but dangerous for security tooling
• What was rejected: Everything. Autonomous agents making security decisions without human review is antithetical to Last Mile's philosophy
• Why rejected: Security findings must be human-reviewed. Auto-fix is suggested, never auto-applied. The "super autonomous" model is inappropriate for a trust product

LangChain

• Language: Python/JavaScript | Risk Level: 🟡 MEDIUM
• What it is: The dominant LLM application framework
• What was taken: Patterns only — chain-of-thought prompting, structured output parsing, tool-use patterns
• What was rejected: All code. Zero imports. No langchain in package.json
• Why rejected: LangChain is a massive dependency tree with frequent breaking changes. For a security product, every dependency is attack surface. The patterns are simple enough to implement directly
• Implementation: Direct API calls to Claude/OpenAI with structured JSON output schemas

---

Memory-First

memU

• Risk Level: 🟡 MEDIUM
• What it is: Memory-first agent architecture with tiered storage
• What was taken: Three-tier memory pattern — hot (KV), warm (D1), cold (Vectorize) storage tiers for different access patterns
• What was rejected: Custom memory management code, Python runtime
• Why relevant: Directly shaped the storage architecture. Scan results go to D1 (queryable), hot config to KV (fast), semantic patterns to Vectorize (searchable)
• Implementation: KV for session/config cache, D1 for structured results, Vectorize for pattern similarity

memU Bot

• Risk Level: 🟠 HIGH
• What it is: Chatbot built on memU memory system
• What was taken: Nothing
• What was rejected: Everything — chatbot interaction model doesn't apply to a scanner
• Why rejected: Last Mile is a CLI/CI tool, not a conversational agent

---

Computer Use

Agent S3

• Risk Level: 🟡 MEDIUM
• What it is: Computer-use agent for browser/desktop automation
• What was taken: Concept deferred to Phase 5 — browser-based security testing (OWASP ZAP-style) via Cloudflare Browser Rendering
• What was rejected: Current implementation — desktop automation is out of scope
• Why deferred: Browser rendering API integration for runtime security testing is a future capability, not a current priority

---

Lightweight

Nanobot

• Stars: ~4k lines | Language: Python | Risk Level: 🟢 LOW
• What it is: Minimal agent loop in under 4,000 lines of Python
• What was taken: Agent loop pattern — the reason → act → observe → repeat cycle that structures how each agent processes findings
• What was rejected: Python code, specific implementations
• Why relevant: Proved that a useful agent loop can be implemented in very few lines. Influenced the design of Last Mile's agent execution cycle
• Implementation: Each agent follows: reason (what to check) → act (run rules) → observe (collect findings) → repeat (next rule)

---

Bonus

OpenClaw Handbook

• Risk Level: 🟢 LOW
• What it is: Documentation and operational guides for the Claw ecosystem
• What was taken: Documentation patterns — how to structure operational documentation for AI systems
• What was rejected: Content (specific to Claw infrastructure)
• Why relevant: Informed this wiki's structure

NoClaw

• Risk Level: ⚪ NONE
• What it is: The meme entry — "what if we just... didn't?"
• What was taken: The philosophy — sometimes the right amount of infrastructure is zero
• What was rejected: N/A (it's a meme)
• Why relevant: Reinforced the decision to use managed APIs instead of self-hosted inference. NoClaw is unironically the correct answer for most teams

---

Summary Matrix

Repo	Risk	Taken	Category
OpenClaw	🔴 CRITICAL	Nothing (replaced by Claude + Workers AI)	Inference
NanoClaw	🔴 CRITICAL	Nothing (replaced by Workers AI)	Inference
PicoClaw	🟠 HIGH	Concept only	Inference
ZeroClaw	🟠 HIGH	Concept only	Inference
MimiClaw	🔴 CRITICAL	Nothing	Inference
OpenFang	🟡 MEDIUM	Agent isolation pattern	Multi-Agent
CrewAI	🟡 MEDIUM	Role-based crew pattern	Multi-Agent
AutoGen	🟢 LOW	Conversational pattern	Multi-Agent
SuperAGI	🟠 HIGH	Concept only (autonomous = dangerous)	Multi-Agent
LangChain	🟡 MEDIUM	Patterns only (zero imports)	Multi-Agent
memU	🟡 MEDIUM	Three-tier memory pattern	Memory
memU Bot	🟠 HIGH	Nothing	Memory
Agent S3	🟡 MEDIUM	Deferred to Phase 5	Computer Use
Nanobot	🟢 LOW	Agent loop pattern	Lightweight
OpenClaw Handbook	🟢 LOW	Documentation patterns	Docs
NoClaw	⚪ NONE	The philosophy	Meme