← Reddit

I built a local-first routing layer for Claude Code that decides which agent handles your request, what it can access, and logs every handoff as an auditable receipt

Reddit · Hot-Leadership-6431 · June 16, 2026
Agentlas Network is a local-first routing layer for AI runtimes that determines which agent handles each request, what resources it can access, and generates auditable receipts of every routing decision. The system uses standardized routing cards shipped by agents and plugins, checking them in a priority order—explicit commands, project overrides, local cards, and Hub fallback—with the Hub receiving only redacted keywords rather than raw prompts. The open-source tool was built with Claude Code and is available free on GitHub.

Detailed Analysis

Agentlas Network, also referred to as Hephaestus Network, is an open-source, local-first routing layer designed to coordinate multiple AI agents, tools, plugins, and memory systems across runtimes including Claude Code, OpenAI Codex, Google Gemini, Cursor, and terminal-based workflows. Built by an independent developer using Claude Code as both the construction tool and the underlying reasoning runtime, the system addresses a structural problem that emerges when AI-assisted development environments scale beyond a single agent: the need for a principled, auditable mechanism to determine which agent handles a given request, what data that agent can access, and how routing decisions are logged and inspectable. The project is published on GitHub under the agentlas-ai organization and is freely available for use.

The core architectural concept is the routing card — a standardized metadata schema that each agent, team, or plugin ships alongside itself, declaring its intended scope, prohibited use cases, required inputs, capabilities, risk profile, memory behavior, and entrypoints. When a user invokes the system via `/hephaestus-network <request>`, the router resolves the destination through a strict priority order: explicit commands first, then project-level routing overrides, then local routing cards, and finally a fallback to the Agentlas Hub. The Hub fallback is designed with a privacy-preserving constraint: only redacted keywords are transmitted, not raw prompts or local project memory. If a Hub-hosted agent is selected, its execution bundle is fetched and run locally within the user's own runtime, preserving local-first semantics even in the fallback case. Every routing decision produces a written receipt that records the selected agent, the rationale, whether the Hub was used, and the local routing state — establishing an auditable chain of custody for agentic decisions.

A companion component, Hephaestus Stormbreaker, handles post-routing execution governance through mechanisms described as scope locks, issue contracts, failure memory, evidence loops, review gates, and final proof. This separation of concerns — routing as one layer, execution governance as another — reflects an architectural philosophy that treats agent orchestration as a multi-phase problem rather than a single dispatch decision. The receipt logging system in particular positions the project within a growing design tradition that treats auditability not as an afterthought but as a first-class system requirement, especially important as agentic workflows increasingly make consequential decisions with limited human-in-the-loop intervention.

The developer's use of Claude Code during construction is itself a notable data point. Claude Code was employed primarily for architecture research and design iteration — shaping the routing card schema, the local-first fallback hierarchy, and the receipt logging format — rather than pure code generation. This reflects an emerging pattern in developer tooling where Claude Code functions as a collaborative design partner capable of engaging at the systems-design level, not merely producing syntactically correct code. The project's reliance on Claude Code as both the build tool and the production reasoning runtime also demonstrates a form of recursive dependency: Claude Code is simultaneously the instrument used to build the system and the substrate on which the system operates.

Agentlas Network enters the market at a moment when multi-agent orchestration is one of the most actively contested problems in applied AI. Frameworks such as LangGraph, CrewAI, AutoGen, and Anthropic's own multi-agent tooling all attempt to address coordination across agents, but most operate at the application layer without providing explicit routing-card contracts or per-decision audit receipts as native primitives. The local-first design philosophy, combined with privacy-preserving Hub fallback and a standardized routing card schema, positions Agentlas Network as an infrastructure-layer approach rather than an application-layer one — a distinction that may prove significant as enterprise and developer-tool use cases demand both flexibility and compliance-grade traceability in agentic systems.

Read original article →