ECOSYSTEM

The Core Stack

Four tools. Each solves a distinct problem. All four converge on the same AI agent — giving it memory, structure, efficiency, and skill at once.

obsidian-forgeKNOWLEDGE CAPTUREalcoveCONTEXT LAYERAI AGENTclaude · cursor · gemini · etc.llm-transpileTOKEN OPTIMIZERepic-harnessORCHESTRATORalcove promoteMCPPostToolUse hook

INTEGRATION

How It Works

01 obsidian-forge + alcove

The Knowledge Pipeline

obsidian-forge runs as a daemon, continuously organizing your Obsidian vault — architecture decisions, research, conventions — and strengthening its knowledge graph. When a note is ready, `alcove promote` imports it into alcove's indexed project docs. From that point, any MCP-compatible agent can query it on demand: BM25 full-text search, vector embeddings, hybrid ranking. Personal notes become authoritative, searchable project context.

02 llm-transpile → agent

The Invisible Optimizer

llm-transpile integrates directly with the agent's file-reading pipeline. Install it once and it hooks into Claude Code's PostToolUse event — every `.md`, `.html`, or `.txt` file an agent reads is automatically compressed into a structured bridge format before it lands in the context window. No prompting required. The agent reads the same information, uses fewer tokens, and reaches further into complex tasks without hitting limits.

03 epic-harness → agent

The Agent That Learns

epic-harness replaces 30+ slash commands with 6 and adds a Ring 0 hook system that observes every tool use, detects failure patterns, and reflects them into a cross-project SQLite knowledge graph. This memory is independent of alcove — it tracks decisions, resolutions, and patterns across sessions, not documents. The result: an agent that improves over time, auto-triggers the right skill for the current context, and can evolve new skills from its own failure history.

Compound Effect

When All Four Converge

In a single agent session all four are active simultaneously. The agent reads a file — llm-transpile compresses it. It needs project background — alcove surfaces the right docs via MCP. It chooses a strategy — epic-harness has seen this pattern before and dispatches the proven skill. The knowledge behind those docs was shaped by obsidian-forge weeks earlier. No single tool produces this result. Together they do.

THE FOUR TOOLS

Each Layer, Explained

obsidian-forge Knowledge Capture
GitHub

Obsidian vault automation daemon. Generates notes, strengthens the knowledge graph, auto-commits, and runs headlessly as a single binary. Feeds alcove via `alcove promote`.

RustCLIProductivity
alcove Context Layer
GitHub

MCP server for on-demand project docs. Stores architecture decisions and conventions; serves them via BM25 + vector hybrid search to any MCP-compatible agent.

RustMCPAI Agents
llm-transpile Token Optimizer
GitHub

Hooks into the agent's PostToolUse event. Every file read is auto-transpiled into a token-efficient bridge format — same fidelity, fewer tokens, no agent configuration needed.

RustCLILLM
epic-harness Orchestrator
GitHub

6 commands replace 30+. Ring 0 hooks observe, reflect, and build a SQLite pattern memory across sessions. Detects failure patterns and evolves new skills automatically.

Claude CodePluginAI Agents