Skip to content
Youssef Alouani

Open source

Tools I built in public, from a Claude Code setup generator to an MCP server for PostgreSQL.

AI Stones

An AI harness for Claude Code: Markdown skills, guardrail hooks and JSON settings that protect and improve how the model operates.

  • Markdown
  • JSON
  • Hook scripts
  • MIT
AI Stones: an AI harness for Claude CodeClaude Code runs inside a harness of hooks (protection: access guard, blocked rm, atomic commits, workflows gate; quality: complexity, naming, format; memory: history and code graph), Markdown skills and JSON settings. ai-stones init detects the stack (languages, frameworks, package manager, commands, conventions), asks for the context it cannot infer, then installs only the selected setup: AGENT.md and CLAUDE.md, .mcp.json, and .claude/ with settings, hooks, skills, commands and tools. stones.json records answers and file hashes so re-runs update unedited files and preserve edited ones.ModelClaude Code sessiontool calls · edits · commandsHooksProtectionaccess · rm · CI gateQualitycomplexity · namingMemoryhistory · code graphKnowledgeSkills24 Markdown guidesSettingsJSON permissions · MCPAGENT.mdproject briefInstallnpx ai-stones initdetect stack · ask · install selectedstones.jsonanswers · file hashes
  • Guardrail hooks: access guard, blocked rm, atomic commits, and a gate that holds merges and publishes until local CI passes.
  • 24 skills that load only when a task needs them, from test-first development to system prompts and golden sets.
  • Installed per project with npx ai-stones init, tailored to the detected stack; MIT on npm.

archio

Object storage in Go with a multi-node, high-availability cluster and semantic search over stored files, built on Docling.

  • Go
  • Apache-2.0
archio object storageClients call a Go JSON REST API running on a multi-node, high-availability cluster. Admin endpoints use HTTP Basic auth; every other endpoint uses a group-scoped API key, except public read URLs for a single object version. Object bytes go to the storage path; metadata, versions and keys live in PostgreSQL with pgvector enabled; Redis is connected and health-checked, reserved for caching; RabbitMQ receives object.version.created events, which drive Docling text extraction; embeddings are stored with pgvector for semantic search.ClientsAdminHTTP BasicAppsgroup API keysPublic read URLsone version eachAPIGo REST APIbuckets · immutable versionsClusterNode 1object bytesNode 2replicaNode nreplicaData and searchPostgreSQLEnt · pgvectorDoclingextraction · embeddingsRedisRabbitMQobject.version.created
  • File processing, storage and delivery optimisation.
  • Multi-node cluster setup for high availability.
  • Semantic search over documents, built on Docling.

translator

Low-cost, context-aware translation: NLLB-200 for simple text, an LLM above a complexity threshold, behind a three-tier cache.

  • Python
  • Apache-2.0
translator: low-cost, context-aware translationA client sends text to a FastAPI service. A three-tier cache on PostgreSQL with pgvector answers first: exact, normalized, then semantic matches. On a miss, the service routes by complexity: simple text goes to NLLB-200 3.3B, and text above the complexity threshold goes to an LLM (Gemini, Ollama or OpenRouter). The LLM can call semantic search as a tool to fetch validated translation examples. multilingual-e5 scores quality, and results are kept as JSON documents in S3 or MinIO.RequestFastAPI servicetranslate · scoreThree-tier cacheExactNormalizedSemanticPostgreSQL · pgvectorRouted by complexityNLLB-200 3.3Bsimple textLLMGemini · Ollama · OpenRouterQuality and storagemultilingual-e5quality scoreS3 / MinIOJSON documentsExample searchvalidated translations, as an LLM tool
  • FastAPI service routing between NLLB-200 3.3B and an LLM (Gemini, Ollama or OpenRouter) by complexity.
  • Exact, normalized and semantic cache on PostgreSQL with pgvector.
  • Semantic search exposed to the LLM as a tool that fetches validated translation examples, lifting domain accuracy even with local LLMs.
  • multilingual-e5 quality scoring; S3/MinIO JSON documents.

Tracking kit

Used in production

A Go/Gin user-recognition service and a TypeScript browser SDK, with every event HMAC-signed.

  • Go (Gin)
  • TypeScript
  • Apache-2.0
Illustration of server-side tracking: a browser window sends events to a secured server and database, which feed analytics charts
Tracking kit: signed events for user recognitionIn the browser, the TypeScript tracking-sdk wraps tracking calls and signs every event with HMAC. The Go service, tracker, built on Gin, receives each event and verifies its signature, so only events from trusted clients are accepted, and then uses them for user recognition.BrowserWebsite or apptracking-sdkTypeScriptEvery eventHMAC signaturetrusted clients onlyServicetrackerGo · Gin · verifies signaturesResultUser recognition
  • Gin service that receives and verifies tracking events for user recognition.
  • TypeScript wrapper for browser-based tracking.
  • HMAC request signing, so the service only accepts events from trusted clients.
  • Contributed to the community as open source, and already running in production.

pg_inspector

An MCP server that gives AI assistants safe, read-only access to PostgreSQL.

  • Python
  • Apache-2.0
Illustration of an MCP server chip next to a magnifying glass labelled Debug and Inspect over streams of queries and logs
pg_inspector: read-only PostgreSQL for AI assistantsAn AI assistant connects to the pg_inspector MCP server over Streamable HTTP with an API key, or over stdio. The server offers read-only SQL, schema inspection, EXPLAIN plans and a health snapshot, plus five guided prompts for common investigations. Every query reaches PostgreSQL read-only.ClientAI assistantMCP clientTransportStreamable HTTPAPI keystdiopg_inspector toolsRead-only SQLSchemaEXPLAINHealthDatabasePostgreSQLread-only access5 guided promptscommon investigations
  • Streamable HTTP with an API key, and stdio.
  • Read-only SQL, schema inspection, EXPLAIN plans and a health snapshot.
  • Five guided prompts for common investigations.