Researched 2026-10-12. Stars and "last push" come from the GitHub REST API on 2026-10-12 (api.github.com/repos/OWNER/REPO). "Latest release" comes from each repo's /releases/latest page (year not shown on page; all read as 2026). Items marked [unverified] come from my background knowledge or a single secondary source and were not confirmed against a primary page this session. License "NOASSERTION" means GitHub could not auto-detect it; check the LICENSE file before depending on it.
| # | Repo | Why, one line |
|---|---|---|
| 1 | PrefectHQ/fastmcp | Fastest path to Ria's MCP server: streamable HTTP, bearer/OAuth auth, stateless mode. |
| 2 | UKGovernmentBEIS/inspect_ai | The de-facto eval framework (UK AISI); run Ria's own regression and red-team evals on it. |
| 3 | NVIDIA/garak | Scanner for probing the Moltbook agents and other LLM endpoints; emits AVID-format reports. |
| 4 | promptfoo/promptfoo | CLI red-team and prompt regression in CI; cheap to wire to cron. |
| 5 | langfuse/langfuse | Self-hosted trace, cost and action log for every Ria run. |
| 6 | BerriAI/litellm | One gateway across Anthropic, OpenRouter, Ollama with per-key budgets and spend logs. |
| 7 | miniflux/v2 | Minimal self-hosted feed reader with an API; backbone of source ingestion. |
| 8 | adbar/trafilatura | Best-in-class article text extraction for the incident and news pipeline. |
| 9 | responsible-ai-collaborative/aiid + weekly snapshots | Seed the incident corpus from the AI Incident Database dumps. |
| 10 | anthropic-experimental/sandbox-runtime | Seatbelt (macOS) / bubblewrap (Linux) wrapper to contain Ria's own tool-running agents on the Mini with no Docker. |
Watch list (monitor, do not adopt yet): assafelovic/gpt-researcher, blazickjp/arxiv-mcp-server, NVIDIA-NeMo/Guardrails, guardrails-ai/guardrails, openai/openai-agents-python.
| Project | URL | License | Stars | Last push / release | What it does | How Ria uses it |
|---|---|---|---|---|---|---|
| NeMo Guardrails | https://github.com/NVIDIA-NeMo/Guardrails | NOASSERTION (README says Apache-2.0 [unverified]) | ~7.3k | push 2026-10-10; release v0.24.1 (16 Sep) | Programmable rails (Colang) for input/output/dialog/retrieval/execution. | Monitor. Borrow the rail taxonomy for Ria's output checks on the daily brief. Heavy for a single-agent pipeline. |
| Guardrails AI | https://github.com/guardrails-ai/guardrails | Apache-2.0 | ~7.5k | push 2026-10-09 | Validators (Hub) plus structured output validation. | Use for schema/PII/toxicity validators on brief output if a lighter check than NeMo is wanted. |
| LLM Guard (Protect AI) | https://github.com/protectai/llm-guard | MIT | ~3.2k | Archived, last push 2026-07-08 | Scanner toolkit for prompt injection, PII, secrets. README says no longer maintained (StationX review). | Do not adopt. Mentioned because many tutorials still cite it. |
| Rebuff | https://github.com/protectai/rebuff | Apache-2.0 | ~1.5k | Archived, last push 2024-08-07 | Prompt-injection detector. | Dead. Skip. |
| Invariant Labs guardrails | https://github.com/invariantlabs-ai/invariant | Apache-2.0 | ~470 | push 2026-01-12 (quiet) | Rule language for agent traces/tool calls. Companion trace viewer: explorer (~64 stars, Apache-2.0). | Study the rule DSL for Ria's containment policy. Low activity; company status [unverified]. |
| Lakera Guard / PINT benchmark | https://github.com/lakeraai/pint-benchmark | MIT | ~200 | Archived, push 2026-04-16 | Commercial runtime guard; Check Point agreed to acquire Lakera for ~$300M (CyberScoop); I did not confirm the close. | Hosted and closed. Not a reuse candidate. |
| Meta PurpleLlama (Llama Guard, Prompt Guard, LlamaFirewall, CyberSecEval) | https://github.com/meta-llama/PurpleLlama | NOASSERTION (mixed; Llama community license on weights) | ~4.4k | push 2026-09-29 | Safety classifiers and the CyberSecEval benchmark suite. | Run Llama Guard / Prompt Guard locally via Ollama or HF as a cheap second-opinion filter on scraped Moltbook content. Check weight license terms. |
| OpenAI Agents SDK guardrails | https://github.com/openai/openai-agents-python | MIT | ~30.0k | push 2026-10-08 | Input, output and tool guardrails with a tripwire exception; input guardrails run parallel or blocking (docs). | Pattern to copy: blocking guardrail before spend, tripwire raises and halts. |
| Claude Code hooks | https://code.claude.com/docs/en/hooks-guide | (doc, product feature) | n/a | n/a | Deterministic shell hooks at lifecycle points (PreToolUse etc.) to allow/deny tool calls. | Ria already runs headless claude -p; put cost-cap and path-allowlist enforcement in a PreToolUse hook, not in the prompt. |
| Project | URL | License | Stars | Last push | What it does | Ria use |
|---|---|---|---|---|---|---|
Anthropic sandbox-runtime (srt) | https://github.com/anthropic-experimental/sandbox-runtime | Apache-2.0 | ~5.5k | n/a (not shown) | OS-level fs and network restriction, no container; JSON policy file; network and writes denied by default (doc). Research preview, config may change. | Primary containment on the Mac Mini. Wrap every Ria worker: srt --settings ria.json <cmd>. |
| E2B | https://github.com/e2b-dev/E2B | Apache-2.0 | ~14.3k | 2026-10-09 | Cloud/self-host Firecracker sandboxes with SDKs. | Use only if Ria must run untrusted code (e.g. exploit repros). Cloud cost applies. |
| Vercel Sandbox SDK | https://github.com/vercel/sandbox | Apache-2.0 | ~210 | 2026-10-12 | Ephemeral Firecracker microVMs hosted by Vercel. | Fits since Byron already deploys on Vercel; for web-hosted risky jobs. |
| Firecracker | https://github.com/firecracker-microvm/firecracker | Apache-2.0 | ~37.3k | 2026-10-09 | microVM monitor under E2B and many others. Linux/KVM only. | Not on the Mac Mini directly. Reference. |
| gVisor | https://github.com/google/gvisor | Apache-2.0 | ~19.6k | 2026-10-12 | User-space kernel for container isolation. Linux. | Use if Ria later moves to a Linux host. |
| Daytona | https://github.com/daytonaio/daytona | GitHub shows none | ~71.6k | API reports archived=true with push 2026-07-24 | Sandbox infra for agents. The archived flag is odd for a 71k-star repo; likely moved or restructured, [unverified]. | Verify before considering. |
| OpenHands | https://github.com/OpenHands/OpenHands (old: All-Hands-AI/OpenHands, now redirects) | MIT per review [unverified] | 60-86k (sources disagree) | n/a | Coding agent; each session runs in a Docker sandbox. | Reference design for a Docker runtime per task. |
| AgentDojo | https://github.com/ethz-spylab/agentdojo | MIT | ~909 | 2026-06-02 | Benchmark/sandbox for prompt-injection attacks and defenses on tool-using agents. | Use the task suites as a red-team corpus for Ria's own agents. |
| Docker/Colima | (already on the Mini) | n/a | n/a | n/a | Container isolation. | Fallback when srt policy is too coarse; combine with read-only mounts. |
| Project | URL | License | Stars | Latest release | Notes / Ria use |
|---|---|---|---|---|---|
| Langfuse | https://github.com/langfuse/langfuse | NOASSERTION (MIT core plus ee/ dirs [unverified]) | ~35.7k | v4.56.0 (09 Oct) | Self-hostable traces, cost, evals, prompt mgmt. Use as Ria's run ledger UI. |
| Arize Phoenix | https://github.com/Arize-ai/phoenix | NOASSERTION (Elastic License 2.0 [unverified], source-available, not OSI) | ~11.8k | push 2026-10-12 | OpenTelemetry/OpenInference native tracing and evals. Alternative to Langfuse; license is more restrictive. |
| OpenTelemetry GenAI semantic conventions | https://github.com/open-telemetry/semantic-conventions (GenAI docs moved to https://github.com/open-telemetry/semantic-conventions-genai) | Apache-2.0 | ~667 | n/a | Standard attribute names (gen_ai.*). Emit these from Ria so any backend works. Stability level not confirmed this session. |
| AgentOps | https://github.com/AgentOps-AI/agentops | MIT | ~5.9k | push 2026-06-25 | Agent session replay. Slower activity; watch only. |
| Helicone | https://github.com/Helicone/helicone | Apache-2.0 | ~6.2k | push 2026-10-10 | LLM proxy with logging and cost. Redundant if LiteLLM is the gateway. |
| LiteLLM | https://github.com/BerriAI/litellm | NOASSERTION (MIT with enterprise dir [unverified]) | ~61.0k | v1.105.0 (11 Oct) | Gateway and per-key/per-team budgets, spend logs, 100+ providers incl. Anthropic, OpenRouter, Ollama. Recommended meter for the cost ledger. |
| Invariant tracing | see (a) | Apache-2.0 | n/a | n/a | Trace format with rule checks. |
| Project | URL | License | Stars | Latest release | Ria use |
|---|---|---|---|---|---|
| Inspect AI | https://github.com/UKGovernmentBEIS/inspect_ai | MIT | ~3.0k | 0.3.278 (09 Oct) | Write evals (agentic, tool use, sandboxed) as tasks; log files; used by AISI. |
| Inspect Evals | https://github.com/UKGovernmentBEIS/inspect_evals | MIT | ~700 | push 2026-10-12 | Community eval catalogue (safety, cyber, agentic). Mine for what to monitor. |
| lm-evaluation-harness | https://github.com/EleutherAI/lm-evaluation-harness | MIT | ~14.2k | push 2026-09-14 | Standard capability benchmarks for local Ollama/HF models. |
| HELM | https://github.com/stanford-crfm/helm | Apache-2.0 | ~2.9k | push 2026-09-01 | Holistic evals and public leaderboards; monitor for new safety scenarios. |
| promptfoo | https://github.com/promptfoo/promptfoo | MIT | ~25.9k | releases tagged per component (latest seen code-scan-action-0.2.2, Oct 8) | Declarative evals plus red-team plugins in YAML. Easiest to cron. |
| garak | https://github.com/NVIDIA/garak | Apache-2.0 | ~9.5k | v0.17.0 (09 Sep) | Vulnerability scanner with probes/detectors; .avid.jsonl export (reference). |
| PyRIT | https://github.com/Azure/PyRIT | MIT | ~120 shown | Archived 2026-03-27 per DEV Community (secondary source) | The API shows only ~120 stars, so a move or reset is possible; I found no confirmation of a successor. Do not depend on it. |
| DeepEval | https://github.com/confident-ai/deepeval | Apache-2.0 | ~18.8k | push 2026-10-12 | Pytest-style LLM unit tests; useful for brief quality checks (faithfulness to sources). |
| Source | URL | Access | License / terms | Ria use |
|---|---|---|---|---|
| AI Incident Database (AIID) | https://incidentdatabase.ai/ | Weekly snapshots at https://incidentdatabase.ai/research/snapshots/ in JSON, MongoDB archive and CSV (~105-107 MB tar.bz2, newest seen 2026-08-10 on cached pages). I found no public GraphQL/REST docs in search, though the site's own app uses an API [unverified]. Code: https://github.com/responsible-ai-collaborative/aiid (~273 stars, push 2026-10-07). | Repo license NOASSERTION; the site asks users to tell them about use cases. Cite McGregor (2021). | Seed corpus; diff each weekly snapshot to detect new incidents; link Ria entries to incident IDs. |
| OECD AI Incidents and Hazards Monitor (AIM) | https://oecd.ai/en/incidents-methodology | Web UI. Fed by Event Registry news clusters, LLM-classified into incident/hazard (OECD methodology). Event Registry data page: https://oecd.ai/en/eventregistry. | OECD terms; no bulk API confirmed. | Copy the pipeline pattern (news clusters then LLM classify then dedupe by embedding similarity). Cross-check Ria's catalogue against it. |
| MIT AI Risk Repository | https://airisk.mit.edu/ | Downloadable spreadsheet database; peer-reviewed version (30 Mar 2026) reports 1,725 risks from 74 frameworks (MIT FutureTech); April 2025 update had 1,612 risks / 65 documents (blog). | Sheet link and license not confirmed here. | Use its causal and domain taxonomy as Ria's tagging scheme. |
| AIAAIC Repository | https://www.aiaaic.org/aiaaic-repository | Public spreadsheet/web. I could not confirm download method or license (commonly CC BY-SA [unverified]). | Verify before ingest. | Second incident source, broader "controversies". |
| AVID | https://avidml.org/ ; code https://github.com/avidml/avid-db (MIT, ~17 stars, push 2026-03-26) | Open data schema; avidtools pip package; garak and Giskard export to it (Giskard docs). Official bulk export not confirmed. | MIT code. | Use the schema for Ria's vulnerability findings so they interoperate. Low activity. |
| MITRE ATLAS | https://github.com/mitre-atlas/atlas-data | ~192 stars, push 2026-09-15 | NOASSERTION | Machine-readable adversary technique taxonomy; tag red-team findings with ATLAS IDs. |
| Project | URL | License | Stars | Last push | Lesson for Ria |
|---|---|---|---|---|---|
| Sakana AI Scientist v2 | https://github.com/SakanaAI/AI-Scientist-v2 | NOASSERTION | ~7.3k | 2025-12-19 | Agentic tree search over experiments. Quiet for 10 months; read, do not run (executes generated code). |
| Agent Laboratory | https://github.com/SamuelSchmidgall/AgentLaboratory | MIT | ~5.9k | 2025-08-20 | Stage pipeline (lit review, plan, run, write). Stale. |
| STORM | https://github.com/stanford-oval/storm | MIT | ~31.6k | 2025-09-30 | Perspective-guided question asking then outline then cited article. Copy the "multi-perspective questions" prompt pattern for Ria's entity profiles. |
| GPT Researcher | https://github.com/assafelovic/gpt-researcher | Apache-2.0 | ~30.0k | 2026-10-10 | Active; planner/executor/publisher with web sources and MCP support. Best reference for Ria's research loop. |
| Open Deep Research (LangChain) | https://github.com/langchain-ai/open_deep_research | MIT | ~12.7k | Archived, push 2026-08-10 | Read for design only. |
| PaperQA2 | https://github.com/Future-House/paper-qa | Apache-2.0 | ~9.4k | 2026-09-25 | RAG over PDFs with citations; use for answering questions over the saved paper corpus. |
| Semantic Scholar API | https://www.semanticscholar.org/product/api | Free; unauthenticated traffic shares a pooled 1000 RPS; an API key starts at 1 RPS (overview, tutorial). Bulk datasets available. | n/a | n/a | Citation counts, related papers, author data. Request a key. |
| arXiv API | https://info.arxiv.org/help/api/tou.html | Terms: no more than one request per 3 seconds, single connection, limits apply across all your machines (ToU). | n/a | n/a | Use OAI-PMH or RSS by category (cs.AI, cs.CL, cs.CY) once a day, cached. |
| Elicit, scite, Connected Papers | https://elicit.com , https://scite.ai , https://www.connectedpapers.com | Closed SaaS | n/a | n/a | Not researched in depth this session; treat as inspiration, not dependencies. |
| Project | URL | License | Stars | Latest | Notes |
|---|---|---|---|---|---|
| Miniflux | https://github.com/miniflux/v2 | Apache-2.0 | ~9.8k | 2.3.3 (24 Jul) | Single Go binary + Postgres, REST API, fetch-full-content option. Ria's feed hub. |
| RSS-Bridge | https://github.com/RSS-Bridge/rss-bridge | Unlicense | ~9.3k | push 2026-10-10 | Generates feeds for sites without them. |
| feedparser | https://github.com/kurtmckee/feedparser | NOASSERTION (BSD-2 [unverified]) | ~2.4k | push 2026-10-06 | Python parsing if not using Miniflux API. |
| trafilatura | https://github.com/adbar/trafilatura | Apache-2.0 | ~7.0k | v2.3.1 (6 Oct) | Main-text extraction plus metadata, sitemaps, feeds. |
| newspaper4k | https://github.com/AndyTheFactory/newspaper4k | MIT | ~1.2k | push 2026-08-24 | Alternative extractor; trafilatura usually wins. |
| Crawl4AI | https://github.com/unclecode/crawl4ai | Apache-2.0 | ~85.3k | push 2026-10-12 | LLM-friendly crawler (Playwright) for JS-heavy pages. |
| yt-dlp | https://github.com/yt-dlp/yt-dlp | Unlicense | ~197k | push 2026-10-10 | Subtitles/audio for talks and podcasts. |
| whisper.cpp | https://github.com/ggml-org/whisper.cpp | MIT | ~54.3k | push 2026-10-12 | Local transcription on Apple Silicon when no captions exist. |
Reddit .json | https://www.reddit.com/dev/api/ | Reports say unauthenticated .json was broadly blocked from about 30 May 2026 (DEV Community); another source says only throttled (redditapis.com). All third-party blogs. | n/a | n/a | Treat as unreliable. Use OAuth app (~100 QPM free tier per those sources) or skip Reddit. |
| Nitter | https://github.com/zedeus/nitter | n/a | n/a | Reports: X sent cease-and-desist 24 Aug 2026, repo archived, project said it would continue 6 Sep, later reports of it going down again (Gigazine, Privacy Guides thread). | Do not depend on Nitter. Use the official X API only if needed, or follow people via their newsletters/blogs. |
SDKs
uv run mcp run server.py --transport streamable-http.mcp.run(transport="http", host=..., port=8000) serves /mcp; mcp.http_app() for uvicorn; stateless_http=True for multi-worker; auth via token verifiers or OAuth providers; mountable inside FastAPI/Starlette (pass mcp_app.lifespan).Existing servers worth reusing (as clients of Ria or as sources)
search_nodes, read_graph, open_nodes), filesystem, git.read_wiki_structure, read_wiki_contents, ask_question, no auth, public repos; docs) and Context7 (https://mcp.context7.com/mcp, streamable HTTP with OAuth, API key for higher limits; overview).Exposure with auth for multiple apps: see mcp-and-cost-design.md.