AI and agents
Agent orchestration, MCP servers, LLM skills and model tooling
5 repositories
QuietaippPython · Generativ AI Prompt extensions | Release v0.1.3 | GitHub | MIT | 0 open issues | 0 open pull requests | 2 findings | Health 70Watch | |
|---|---|---|---|---|---|---|---|---|
ActivededupdRust · Always-on local daemon that deduplicates Hugging Face models across Ollama, LM Studio, MLX and llama.cpp using one content-addressed store, and archives them for offline use. | Release v0.1.0 | GitHub | MIT | 1 open issues | 0 open pull requests | 0 | Health 100Good | |
Quietllm-skillsLLM Skills catalog | none | – | APL2 | 0 open issues | 0 open pull requests | 1 finding | Health 90Good | |
Quietmcp-pkg-localTypeScript · MCP server providing LLMs direct access to local package source code to eliminate API hallucinations | 0.1.1 | npm | MIT | 1 open issues | 4 open pull requests | 3 findings | Health 65Watch | |
Quietmcp-web-scraperTypeScript · An intelligent MCP Web Scraper that support Consent Dialog Handling and MCP Playwright | Release v1.0.0 | GitHub | MIT | 3 open issues | 5 open pull requests | 3 findings | Health 65Watch |
Health
Health 82Good5 repositories weighed: 0 at attention, 3 to watch
- Unreleased, medium, costs 20 points: aipp: Unreleased changes since v0.1.3, released 20 months ago
- Unreleased, medium, costs 20 points: mcp-pkg-local: Unreleased changes since 0.1.1, released 15 months ago
- Unreleased, medium, costs 20 points: mcp-web-scraper: Unreleased changes since v1.0.0, released 17 months ago
- Quiet, low, costs 10 points: aipp: No change for 20 months
- Quiet, low, costs 10 points: llm-skills: No change for 6 months
- Quiet, low, costs 10 points: mcp-pkg-local: No change by a person for 14 months
- Quiet, low, costs 10 points: mcp-web-scraper: No change by a person for 17 months
- Old updates, low, costs 5 points: mcp-pkg-local: 4 dependency updates open for more than 30 days
- Old updates, low, costs 5 points: mcp-web-scraper: 5 dependency updates open for more than 30 days
Pull requests
8 waiting for a reviewchore: update @types/uuid to 11.0.0
Needs review, waits for a reviewmcp-web-scraper #51renovate[bot] (automation)
chore: update typescript to 5.9.3
Needs review, waits for a reviewmcp-pkg-local #18renovate[bot] (automation)
chore: update @modelcontextprotocol/sdk to 1.32.2
Needs review, waits for a reviewmcp-pkg-local #10renovate[bot] (automation)
chore: update typescript-eslint monorepo to 8.72.2
Needs review, waits for a reviewmcp-pkg-local #23renovate[bot] (automation)
chore: update @types/better-sqlite3 to 9.6.0
Needs review, waits for a reviewmcp-pkg-local #54renovate[bot] (automation)
Issues
5 needing triageAfter compiling mac M1 tsc-alias, the file cannot be found
Needs triage, waits for a maintainer to look at itmcp-web-scraper #40packaaaidle
Needs triage, waits for a maintainer to look at itmcp-web-scraper #52kalle07idle
Needs triage, waits for a maintainer to look at itdedupd #1renovate[bot] (automation)idle
Needs triage, waits for a maintainer to look at itmcp-pkg-local #1renovate[bot] (automation)idle
Needs triage, waits for a maintainer to look at itmcp-web-scraper #1renovate[bot] (automation)idle
Releases
Up to date- descoped/dedupdv0.1.0dedupd 0.1.0 by oranheim
Release notes
The first release of dedupd: the daemon, the
dedupdcommand and the macOS app, in one disk image signed with Developer ID and notarised by Apple.dedupd keeps each model file once, in a content-addressed store, and gives Ollama, Bionic and LM Studio, and the Hugging Face cache (
hf,huggingface_hub,mlx_lm, llama.cpp) a copy-on-write clone of it. See the README and the documentation.Requirements
- macOS 26 on Apple silicon.
- APFS, with the dedupd store (
~/.dedupd) and the apps' model folders on the same volume.
Install
- Open
dedupd-0.1.0.dmgand drag Dedupd to Applications. - Open Dedupd. It runs in the menu bar.
- Install the background service with the
dedupdcommand the app carries, which copies itself to~/.dedupd/binand registers a launchd agent:
The Install button on the app's Daemon screen runs the same command; it has been tested only against a simulated agent so far./Applications/Dedupd.app/Contents/Helpers/dedupd install
The daemon takes the models the apps already hold into the store; the apps' files are not changed. Quitting the app leaves the daemon running.
Included
- Download once for several apps, adoption of existing models, and replacement of duplicate copies by clones.
- The Hugging Face Hub API and the Ollama registry API on
127.0.0.1:8642, answered from the store. - App profiles: which app takes which format, by which method and at what cost in disk space. 4-bit and 8-bit MLX models are placed into Ollama as clones with a rewritten index.
- An archive disk: a second copy of every stored file, and archive and restore per model.
- The app: Library, Discover, Downloads, Storage and Daemon screens, and the menu bar panel.
Not included yet
- Updates for the app. A new version is installed by replacing the app; the Daemon screen then offers to update the daemon.
- Linux, ModelScope and the catalogue of model families; see the roadmap.
Checksum
92c7a5c1c1d0a0e787ad2deab50602ffd167bf0681e88e940d0888e9f5919872 dedupd-0.1.0.dmg - descoped/mcp-pkg-localv0.1.1Release v0.1.1 - Performance & SQLite Optimization by oranheim
Release notes
🚀 Major Performance Improvements
High-Performance SQLite Cache
- 40x faster validity checks (0.03ms vs 1.2ms)
- WAL mode for optimized concurrent access
- Prepared statements for efficient queries
- MessagePack serialization for binary data storage
- Response times: ~150ms scan, ~10ms read, ~5ms cache hits
Token Efficiency Optimizations
- 90% token reduction for scan-packages (20K → 2K tokens)
- 99% reduction with summary mode (20K → 200 tokens)
- 94% reduction with lazy loading (5K → 300 tokens)
- Smart limits: Default 50 packages for optimal LLM performance
✨ New Features
Advanced Filtering System
- 📊 Summary Mode: Ultra-efficient package counts only
- 🔍 Regex Filtering: Pattern-based package selection
- 📦 Category Filtering: Production vs development dependencies
- 🏷️ Group Filtering: Pre-defined groups (testing, linting, building, etc.)
- 🚫 Type Exclusion: Optionally exclude @types packages
Enhanced Package Intelligence
- 📈 Relevance Scoring: Prioritizes direct dependencies over transitive
- 🎯 Smart Prioritization: Project dependencies ranked higher
- 🌲 Lazy Loading: File trees loaded on-demand for faster navigation
- 🔄 Package Scoring: Popular packages get higher relevance scores
Developer Experience
- 🛠️ TypeScript 5.9+: Strict mode with full type safety
- 📦 ES Modules: Modern import maps (#utils, #tools, etc.)
- 🧪 Comprehensive Testing: 59 tests with 100% pass rate
- 🔧 Better Error Handling: Actionable error messages with suggestions
🏗️ Architecture Improvements
Simplified Caching
- Removed 350+ lines of dead JSON fallback code
- SQLite-only implementation (no unnecessary fallbacks)
- 87% code reduction in cache.ts (407 → 53 lines)
- Synchronous operations for better performance
Test Infrastructure
- Sequential test execution to prevent SQLite database locking
- Directory-based cache naming for proper test isolation
- 100% reliable test suite (54/54 tests pass consistently)
- 60-second timeout for performance-heavy operations
Code Quality
- 0 ESLint errors/warnings
- 0 TypeScript errors
- Safer null handling with nullish coalescing (??)
- Proper type definitions eliminating any types
📊 Statistics
Metric Value Files changed 53 Insertions +8,140 Deletions -2,464 Net change +5,676 lines Test coverage 59 tests (100% pass rate) Dependencies 420 packages (0 vulnerabilities) Build time ~2s (clean environment) 🐛 Bug Fixes
- Fixed SQLite database locking in concurrent test execution
- Resolved test robustness with better package detection
- Fixed TypeScript path mapping for #utils/sqlite-cache
- Improved cache sharing between scan and read operations
- Enhanced error handling and validation across all tools
🔧 Technical Changes
Cache System
- UnifiedCache now SQLite-only (removed PartitionedCache, IndexCache)
- Cache stored in .pkg-local-cache/cache.db (or test-specific variants)
- Automatic WAL mode for concurrent access
- Directory-based naming for test isolation
- 1-hour TTL with smart invalidation
Test Infrastructure
- vitest.config.ts: Sequential execution with pool: 'forks', singleFork: true
- Test timeout: Increased to 60s for performance benchmarks
- Cache isolation: Directory-based DB naming prevents conflicts
- 100% reliability: No more SQLite locking errors
API Consistency
- All cache methods now synchronous
- Improved type safety across all modules
- Better error messages with actionable suggestions
- Consistent logging and debugging output
🏃♂️ Performance Benchmarks
Operation v0.1.0 v0.1.1 Improvement Cache validity check 1.2ms 0.03ms 40x faster Full scan tokens 20,000 2,000 90% reduction Summary mode tokens N/A 200 99% reduction Lazy loading tokens 5,000 300 94% reduction Cache hit response 50ms 5ms 10x faster Test suite reliability 85% 100% Consistent
Full Changelog: v0.1.0...v0.1.1
- descoped/mcp-web-scraperv1.0.0Release v1.0.0 by oranheim
Release notes
Features
🍪 Advanced Cookie Consent Handling
- 30+ Languages: Norwegian, English, German, French, Spanish, Italian, and more
- 25+ Frameworks: OneTrust, Quantcast, Cookiebot, TrustArc, and others
- 100% Success Rate: Tested on major European and American news sites
- Sub-second Performance: <1000ms average consent handling
🔌 MCP Protocol Compliance
- Perfect 10/10 Score: Full MCP 2024-11-05 specification compliance
- Real-time Progress: 5-stage workflow tracking with SSE notifications
- Content Streaming: Live content delivery during extraction
- TypeScript Native: Complete type safety and IntelliSense support
🏆 Complete Browser Automation (29 Tools)
- 100% Microsoft Playwright MCP Parity: All 29 tools implemented
- Core Interactions: Navigation, clicking, typing, form handling, dialogs
- Advanced Features: PDF generation, console monitoring, accessibility testing
- Session Management: Tab management, history navigation, network monitoring
- AI-Powered Vision: Element finding, page annotation, JavaScript execution
- Session Persistence: Maintain browser state across tool calls
- Cookie Consent: Superior handling that Microsoft's implementation lacks
- descoped/aippv0.1.3Release v0.1.3 by oranheim
Release notes
Implemented PromptBuilder with composable TextProcessors.
Created a modular system for AI prompt engineering with a fluent interface.
Key components:
- PromptBuilder with method chaining for prompt construction
- TextProcessor protocol for composable transformations
- Built-in processors for common patterns: variables, JSON, step-by-step, chain-of-thought, few-shot examples, structured output, etc.
- Support for custom processors via lambda functions
Latest commits
Up to dateDrop the Licence section from the site's home page; the footer names the licence0 additions and 5 deletions in 1 file
descoped/dedupdoranheim ecb787c
Drop the note on model licences from the READMEs0 additions and 4 deletions in 2 files
descoped/dedupdoranheim b06cfb3
Drop the affiliation line from the READMEs; keep the note on model licences2 additions and 4 deletions in 2 files
descoped/dedupdoranheim 957f5ac
Remember when Ollama refuses a model, and say so when a job places nothing426 additions and 19 deletions in 11 files
descoped/dedupdoranheim a9e70cc
Give the app an icon made from the monogram64 additions and 130 deletions in 18 files
descoped/dedupdoranheim 4f32dd8
Count the site's visitors with Plausible11 additions and 0 deletions in 1 file
descoped/dedupdoranheim 8cf273b
Lay the site out for phones142 additions and 5 deletions in 3 files
descoped/dedupdoranheim d4878e7
Make the download button download the latest disk image92 additions and 9 deletions in 4 files
descoped/dedupdoranheim fb3e184
Give the site a download button and a footer81 additions and 2 deletions in 3 files
descoped/dedupdoranheim c0f358e
Set the download link while the page is rendered55 additions and 37 deletions in 6 files
descoped/dedupdoranheim 1d609dd