— / 05
Sources — what this review stands on
Archivo fuenteLecciones
Checked on 11 October 2026. For each entry: what I took from it and where it appears in the review. Primary sources (documentation, code, release notes, GitHub issues, model cards) take precedence; I mark blogs and marketing material (blog) or (marketing). Each profile in 01-PROFILES.md also has its own source links.
⚠️ Everything here goes stale within weeks. Treat a number from this page as someone else's dated measurement — to be checked on your own before you base a decision on it.
1. Numbers from GitHub
- GitHub API —
GET /repos/{owner}/{repo},GET /repos/{…}/releases/latest,GET /repos/{…}/security-advisories, called on 11 October 2026 for 218 repositories. → stars, licences, dates of the latest releases and commits, the "archived" status in README.md and the profiles; the number of OpenClaw security advisories (722 in 2026: 14 critical, 249 high, 390 medium, 69 low) and Open WebUI (158 in 2026). releases/latestskips pre-releases, and in monorepos it returns a release of a different package. For LangGraph (thelanggraphpackage), Strands Agents (harness), Kilo Code (JetBrains plugin), DeepSeek Harness and LibreChat I took the version from the project's release pages.
2. Top harnesses
| Harness | Sources | What I took |
|---|---|---|
| LangGraph | models, durable execution, multi-agent patterns, #40219, #35514 | OpenAI-compatible configuration, durability modes, five patterns, reasoning_content and streaming bugs |
| Goose | environment variables, subagents, tool shim, Adversary, move to AAIF, #11142 | subagent model, recipes and cron, security modes, the foundation, known bugs with local providers |
| OpenHands | local models, file-based agents, OpenAI gateway, Canvas v1.25.0, #16459 | connecting via LiteLLM, model: inherit, an agent as a "model", voice input, the fixed concurrency lock |
| Hermes Agent | delegation, providers, configuration, voice, cron, releases | delegation.provider/model/base_url, 10 children by default, context_length, STT/TTS, cron without an agent, the pace of change |
| Pydantic AI | PR #4149 (merged 4 February 2026), OpenAI-compatible APIs, subagents, release v2.0.0, #4878 | ConcurrencyLimiter / ConcurrencyLimitedModel (checked in the code: pydantic_ai/concurrency.py), capabilities v2, heuristics for llama.cpp |
| Deep Agents | overview, subagents, async subagents, Open Deep Research — archive | model per subagent, task in parallel, the successor to Open Deep Research |
| Strands Agents | llama.cpp provider, harness, BidiAgent GA, #4620 | native LlamaCppModel, a ready-made harness, cloud-only voice, 200k context by default |
| OpenCode | providers, agents, server, #7410, #52462 | @ai-sdk/openai-compatible configuration, model and permissions per agent, the Pro/Max login block (9 January 2026), the function.name: null bug |
| DeerFlow | README, config.example.yaml, #1055 | llama.cpp example, model per subagent, scheduled tasks, problems with local models |
| ZeroClaw | providers, delegation, multiple models, #10396 | llamacpp slot, native_tools, delegation forbidden by default, risk profiles |
| Microsoft Agent Framework | OpenAI provider, checkpoints, orchestrations, #8382 | the OpenAIChatClient = Responses API trap, orchestrations, the reasoning_content fix (September 2026) |
| DeepSeek Harness | README, providers, SAFETY.md | compat profile, Claude Code / Codex / ACP subagents, read-only by default, pre-release status |
| Pi | agent README, llama.cpp, models, security | llama.cpp router mode, models.json, virtual models, no approvals by design |
| gptme | providers, tool formats, subagents | OPENAI_BASE_URL, --tool-format xml, subagent isolation modes |
| Moltis | providers, agent presets, voice, local LLM | presets with a model, delegate_only, local STT/TTS, no ROCm path in the built-in backend; number of commits — contributors API |
| QwenPaw | models, multi-agent, channels, #7979 | chat.completions mode, model per agent, the list of channels, the context window bug |
| OpenAI Agents SDK | models, release v0.23.0, repository | OpenAIChatCompletionsModel, disabling tracing, Claude via LiteLLM, voice pipeline |
| n8n | licence, AI Agent Tool, human in the loop, concurrency, security advisories, ACSC alert, #26639, OpenAI Chat Model node | licence terms, model per node, approvals, a limit that does not cover sub-workflows, CVE-2026-21858 (checked in the API: critical, 7 January 2026; per ACSC unauthenticated remote code execution), the "Use Responses API" switch (Chat Completions by default) |
| LibreChat | agents, release v0.8.8, config v1.3.15, releases | endpoints.custom, includeReasoningContent, subagent teams, approval rules, Scheduled Chats |
| Kilo Code | custom subagents, CLI, licence, #12773 | model per subagent, OpenCode core, schema incompatibility with the llama.cpp grammar |
3. Verified separately in the primary source
I checked the claims that changed the ranking the most myself, outside the research agents:
| Claim | Source | Score |
|---|---|---|
| compromise of LiteLLM 1.82.7 and 1.82.8 | BerriAI/litellm#24518 | confirmed, 24 March 2026 |
| Codex CLI removed Chat Completions | openai/codex — discussion #7782 | confirmed: an error with wire_api = "chat" since February 2026 |
| a request for local models for Claude Code subagents | anthropics/claude-code#38698 | open |
| Flowise and Roo Code archived | GitHub API (archived: true) |
confirmed |
| OpenBB under a new organization, licence | GitHub API (OpenBB-finance/OpenBB → openbq-org/OpenBB) |
redirect confirmed; licence change to Apache-2.0 in V5 per the release notes |
| attack on Mastra npm packages | BankInfoSecurity (press) | 140+ packages, attributed to BlueNoroff by Microsoft |
| DeepSeek-V4-Flash: size, licence, no image input | model card | 284B / 13B active, MIT, 1M context, no vision |
| Claude prices | Anthropic pricing | Opus 5.5 4 / 20, Sonnet 5.5 2 / 10, Haiku 5.5 0.10 / 0.50 USD per million tokens; Batch −50% |
| OpenClaw: hundreds of malicious skills, exposed instances | insiderllm (blog), security advisories, security documentation | number of advisories confirmed in the API; figures from industry reports (SecurityScorecard, Koi Security) from a secondary source only |
4. Building blocks
| Group | Key sources | What I took |
|---|---|---|
| gateways | llama-swap, Bifrost — performance and concurrency, Datadog Security Labs on LiteLLM | concurrency with a queue, the course of the compromise |
| search | SearXNG — settings.yml, Crawl4AI — CHANGELOG, Trafilatura, Scrapling, Playwright MCP | JSON disabled by default, Docker server security fixes, the recommendation of CLI over MCP |
| deep research | Local Deep Research, #6820, GPT Researcher | topic subscriptions, no subagent limit, BM25 instead of embeddings |
| prices and changes | changedetection.io, Huginn, urlwatch | price and availability mode, AI rules via llama.cpp |
| finance | OpenBB ODP — MCP, edgartools, CCXT, TradingAgents v0.6.0, yfinance | provider per tier (Opus deep_think), SEC limits, data terms |
| notifications | ntfy — configuration, Apprise — releases, Telegram Bot API — changes | upstream-base-url, breaking v2.0.0, new voice input options in Bot API 10.1–10.3 |
| scheduling | Prefect — server, DBOS, Temporal, systemd timers | SQLite without Docker, queues with a limit, Persistent=true |
| memory | Cognee — LLM providers, Cognee v1.6.0, Graphiti, Mem0 | mode without a cloud key, removal of graph memory from Mem0 OSS |
| observability | Phoenix — self-hosting, Langfuse, OpenLIT, promptfoo | ELv2 licence, AMD GPU monitoring, the acquisition of promptfoo by OpenAI |
| sandboxes | sandbox-runtime, microsandbox, E2B runtime, Daytona | no Docker, micro-VMs, the closing of Daytona's core |
| MCP for DevOps | kubernetes-mcp-server — configuration, mcp-grafana — releases, github-mcp-server, GitLab MCP | read-only modes, telemetry on by default in mcp-grafana 2.0 |
5. Voice and vision
| Topic | Sources | What I took |
|---|---|---|
| Pipecat and Polish | Pipecat, Smart Turn, LiveKit Agents | Smart Turn supports Polish (94.47% per the v3.0 announcement (blog)), LiveKit detectors do not |
| STT in Polish | whisper.cpp, Parakeet-TDT-0.6B-v3, Canary-1B-v2, Qwen3-ASR | WER on FLEURS-pl: Whisper large-v3 4.74%, Canary 6.64%, Parakeet 7.31% (NVIDIA's table) |
| TTS in Polish | Piper, Chatterbox, VoxCPM, Kokoro-82M, F5-TTS | which ones have Polish and under what licence; Chatterbox's watermark; F5 requires ROCm 7.x on RDNA4 |
| speech-to-speech | Moshi, Unmute | no Polish, no ROCm path |
| DeepSeek and vision | DeepSeek-V4-Flash-Vision-Exp, llama.cpp b10766, #28939, multimodal in llama.cpp | the variant with vision, the minimum llama.cpp version |
| vision models and OCR | Qwen3.6-35B-A3B, Qwen3.8-27B, Holo4-35B-A3B, Docling, PaddleOCR-VL-1.5 | sizes, licences, llama.cpp support |
| image generation | ComfyUI, stable-diffusion.cpp, Qwen-Image-Edit-2511, Z-Image-Turbo, FLUX.2 klein 4B | models under Apache-2.0 vs non-commercial |
6. Hybrid, routing and costs
| Topic | Sources | What I took |
|---|---|---|
| planner + executor in coding | Aider: architect/editor | 80.5% and 78.9% with DeepSeek as the editor |
| the cloud splits, the local model processes | Minions — arXiv 2502.15964, repository | 5.7× lower cost at 97.9% of the quality |
| device / cloud routing | Hera — arXiv 2605.24598 | 92.5% effectiveness with the cloud in 46.3% of steps |
| cost of local models alone | OpenJarvis — arXiv 2605.17172 | −25 to −39 pp after swapping Opus for a small model |
| advisor tool | The advisor strategy (vendor) | +2.7 pp at 11.9% lower cost; Claude models only |
| Claude Code and local models | Claude Code subagents, LLM gateways, custom tools in the Agent SDK | no local model support for subagents; the "local agent as a tool" pattern |
| MCP servers delegating to local models | mcp-llama-swap, PAL | Claude stays the orchestrator |
| Goose lead/worker | multi-model in Goose | mode removed, explicit delegation remained |
| DeepSeek in llama.cpp | #29008 — prompt cache in loops, #26398, #26269, b10217 | DSML call parser, cache fixes |
| prices | Anthropic pricing, DeepSeek pricing | the table in 03-ARCHITECTURE.md |
7. Measurements of this machine
| Source | What I took |
|---|---|
| local-llm-bench-results — PARALLEL.md | 22.6 tok/s for 1 agent, 13.6 tok/s per agent with 4 (54.3 in total); prompt reading 1,186 tok/s (1 agent) and 1,694 tok/s in total (4); optimum of 3 agents; 44 s of waiting with 4 agents at 16k tokens each; gpt-oss-20b 141 and 81 tok/s |
| local-llm-bench-results — RESULTS.md | hard task set: DeepSeek-V4-Flash 92 / 93, gpt-oss-20b 47 / 52 |
| this machine's llama-swap configuration | UD-IQ3_XXS quantization (104.2 GB), 512k context across 4 slots, at least 1.4 GB of free VRAM per card, card power limit of 250 W, the gpt-oss-20b-cpu variant |