Detailed Analysis
Hermes, a GitHub-based agentic AI framework, reached 40,000 stars in 46 days — faster than any comparable system in its category, including OpenClaw, which achieved the same milestone in 61 days. Rather than installing Hermes directly, the author of this technical walkthrough chose to reverse-engineer its most valuable components and rebuild them natively within Claude Code, Anthropic's terminal-based coding agent. The core features transplanted include an identity layer (via user.md and memory.md files), shared brand context injection, and multi-client architecture — areas where the author argues Hermes falls short for agency or multi-brand use cases. The self-learning loop that Hermes is celebrated for was deliberately excluded, primarily because the same model that generates skills also validates them, creating a circular self-assessment problem with no external guardrails, version control, or audit trail.
The article's central argument rests on three identified costs of adopting off-the-shelf agentic systems. First, inherited architectural assumptions — Hermes was designed for single-user, single-client deployments, meaning agencies or multi-brand operators would need separate installations per client, fragmenting skills and memory across isolated environments. Second, the opacity problem: when vulnerabilities emerge in systems the user doesn't understand, debugging becomes intractable. OpenClaw, one generational cycle ahead of Hermes, has accumulated over 200 filed vulnerabilities since February, including a security researcher's discovery of 386 malicious packages uploaded by a single threat actor to its skills marketplace. Third, scalability limitations: a non-technical CEO cited in the article spent over 100 hours and $1,000 testing OpenClaw over two months before concluding its bug and security profile made it unfit for production use, subsequently migrating to Claude and replicating roughly 30% of OpenClaw's functionality independently.
The decision to build inside Claude Code rather than on top of Hermes reflects a broader practitioner-level tension emerging across the agentic AI landscape — the tradeoff between rapid deployment using pre-packaged frameworks and the long-term maintainability of custom-built systems. Claude Code, Anthropic's locally-running CLI agent, provides a substrate where developers can construct modular memory, identity, and workflow layers that are portable across projects and swappable as requirements evolve. The author's architecture explicitly addresses multi-tenancy — a gap in Hermes — by injecting client-specific brand voice, ICP (ideal customer profile), and visual identity as discrete context modules rather than baking them into a monolithic identity file.
This development sits within a rapidly accelerating wave of "agentic scaffolding" tooling, where third-party projects attempt to layer persistent memory, self-improvement loops, and identity management onto foundation models. The explosive GitHub adoption rates of both Hermes and OpenClaw signal genuine developer demand for these capabilities, but also underscore that the ecosystem is still maturing. Security vulnerabilities at the skills-marketplace layer — where community-contributed code modules can be injected into agent workflows — represent a systemic risk that centralized, off-the-shelf frameworks are poorly positioned to address at the individual user level. By contrast, rebuilding these systems inside a controlled environment like Claude Code allows practitioners to audit every layer and avoid inheriting unknown threat surfaces.
Anthropic's Claude Code emerges from this analysis not merely as a coding assistant but as a flexible agentic substrate capable of hosting sophisticated memory architectures, multi-client identity systems, and procedural workflow layers — all without the dependency overhead of external frameworks. The author's conclusion that a custom setup is "faster to scale" than Hermes, despite being slower to start, captures a recurring pattern in enterprise software adoption: initial velocity often comes at the cost of long-term flexibility. As agentic systems become more deeply integrated into business operations, the ability to understand, audit, and modify every architectural layer is likely to become a meaningful competitive differentiator for both individual practitioners and organizations deploying AI at scale.
Read original article →