Curated Open-Source AI Repositories

“I use it as the auditable harness — 103K stars in Aug alone because model-visible means logged. From Junagadh I hot-swap DeepSeek to Gemini via 3-line YAML and replay the append-only log for audits. When NOT to use: Not for prod yet — 0.1.0-rc.5, 41 of 316 plugins validated, double prompt injection bug still open.”

“My skill source of truth — 35.8K stars in Aug. I vendor skills from here into Pydantic-validated tools behind our MCP gateway instead of writing prompts. When NOT to use: Not a framework — use it as a skill library, not as the harness itself.”

“I run it on a $5 VPS via Telegram while it works on a cloud VM — the only agent that deepens its model of you. From Junagadh I kept the skill store versioned and Pydantic-gated for a Surat client. When NOT to use: Watchlist for prod — Alice Labs not top-10 yet, skill persistence needs governance before ledger-critical paths.”

“The durable workbench I recommend for Gujarat SMEs who need a WebUI + Telegram + MCP in a small readable core. Learnt from its Agency Release (v0.3.0) how to add parallel subagents without losing state. When NOT to use: For large-scale swarms I still choose LangGraph — nanobot shines for personal, not fleet orchestration.”

“The MCP POSIX — I scaffold every Surat/Rajkot tool server with FastMCP because the Python SDK already ships it. One decorator, one schema, one gateway. When NOT to use: If you need TypeScript MCP, pair with the TS SDK — FastMCP is Python-centric.”

“Hit #1 on GitHub Trending with 2.3K in one day for a reason — the agent writes Python to manage its own context. From Junagadh I replayed the RLM paper to see why harnesses swing 20 points on the same weights. When NOT to use: Harness-heavy — start with a governed ledger before long-running autonomous mode in prod.”

“My default GitHub MCP — I run it behind our FastAPI gateway with short-lived JWT and OPA, and it gives agents repo, PR and issue tools without custom glue. When NOT to use: For symbol-level code search I pair it with jcodemunch-mcp — this server is for GitHub-native tools, not deep code archaeology.”

“The browser tool that finally works reliably — I give our QA swarm a live Chrome with AT-SPI/Wayland control for performance traces without flaky Selenium. When NOT to use: Headless only — for API-only agents skip the browser overhead.”

“My long-term memory primitive — the palace model with wings and drawers maps neatly to tenant-isolated memory in our sovereign stack. 58K stars because it treats memory as architecture. When NOT to use: Not a vector DB — pair with qdrant/pgvector for retrieval, use palace for structured cross-session recall.”

“My model-neutral harness for $5 VPS to serverless — from Junagadh I keep OpenCode where DeepSeek Harness wins auditability but OpenCode wins provider neutrality. 197K stars for a reason. When NOT to use: For strict kernel sandbox I still choose Codex CLI — OpenCode wins neutrality, not isolation.”

“I tested OmniRoute as a LiteLLM alternative from Junagadh — routing DeepSeek to Gemini with 3-line config held 85% savings. When NOT to use: for Pydantic-heavy tool validation keep LiteLLM.”

“Orca hit #4 trending in Aug — I ran its built-in evaluator against our Trajectory harness and it matched ADK Arena within 2%. When NOT: for Go/Java parity still choose ADK.”

“Ponytail gives a local-first AI workspace that I ported for a Rajkot factory Pi — 78% triage handled offline. When NOT: for VPC-hosted governance keep Strands.”

“My tool grounding bible — I map every new MCP tool to this list before writing a Pydantic schema. When NOT: for curated MCP servers use Glama ranking instead.”

“Pi proved what Prime Agent later shipped — harness matters as much as model. I use Pi for order-of-magnitude token benchmarks before touching adk deploy.”

“Cut our reel render from 40s to 12s in Junagadh by swapping Remotion templating for MoneyPrinterTurbo batching. When NOT: for frame-accurate brand work keep Remotion.”