{ "markdown": "

\n agentic-research-engine-oss\n

\n\n

\n \"License\"\n \"PyPI\"\n \"Version\"\n \"Default\"\n \"Tests\"\n \"Interfaces\"\n \"MCP\"\n

\n\n**The best $0 research agent that runs on a laptop.** Open-source\nend-to-end, reproducible, privacy-preserving. No cloud dependency by\ndefault; no telemetry; every LLM call, every source, and every\nverification decision is visible.\n\n---\n\n## Table of contents\n\n- [TL;DR](#tldr)\n- [Why use this instead of…](#why-use-this-instead-of)\n- [Quickstart — Mac local](#quickstart--mac-local)\n- [Quickstart — no install (Google Colab)](#quickstart--no-install-google-colab)\n- [Three ways to drive it](#three-ways-to-drive-it)\n- [What ships](#what-ships)\n- [Domain presets](#domain-presets)\n- [Bring your own documents](#bring-your-own-documents)\n- [MCP + Claude plugin](#mcp--claude-plugin)\n- [Plugin / skill loader](#plugin--skill-loader)\n- [Architecture at a glance](#architecture-at-a-glance)\n- [Repo layout](#repo-layout)\n- [Configuration (env vars)](#configuration-env-vars)\n- [Testing](#testing)\n- [Troubleshooting](#troubleshooting)\n- [Honest limits](#honest-limits)\n- [Status + roadmap](#status--roadmap)\n- [Contributing](#contributing)\n- [License](#license)\n\n---\n\n## TL;DR\n\n**Local-first research agent that verifies its own answers.** Runs on\nGemma 3 4B + Ollama (3.3 GB on disk) for `$0/query`; swaps to any\nOpenAI-compatible endpoint with one env var.\n\n```bash\npip install agentic-research-engine\nagentic-research ask \"what is Anthropic's contextual retrieval?\" --domain papers\n```\n\n| | |\n|---|---|\n| **Interfaces** | CLI · Textual TUI · FastAPI web GUI · MCP server (Claude Desktop / Cursor / Continue) |\n| **Pipeline** | 8-node LangGraph (`classify → plan → search → retrieve → fetch → compress → synthesize → verify`); every node env-toggleable for ablation |\n| **Retrieval** | SearXNG meta-search + trafilatura fetch + hybrid BM25 / dense / RRF; opt-in `bge-reranker-v2-m3` cross-encoder |\n| **Reasoning** | HyDE query expansion · FLARE active retrieval · Chain-of-Verification (Dhuliawala et al 2023) · ThinkPRM step critic |\n| **Domains** | 6 presets (`general` · `medical` · `papers` · `financial` · `stock_trading` · `personal_docs`) — write your own in 10 lines of YAML |\n| **Plugins** | load Claude plugins or `agentskills.io` skills from GitHub or local paths |\n| **Memory** | opt-in local SQLite trajectory log with semantic retrieval; wipe anytime; no telemetry |\n| **Providers** | OpenAI · Groq · vLLM · SGLang · Together · Ollama — any OpenAI-compatible endpoint via `OPENAI_BASE_URL` |\n| **Quality** | 137 mocked tests, zero-network · honest live benchmarks published in [`RESULTS.md`](engine/benchmarks/RESULTS.md) · MIT end-to-end |\n\n---\n\n## Why use this instead of…\n\n| you currently use | we give you |\n|---|---|\n| **Perplexity / ChatGPT Deep Research / Kagi Assistant** | the same reasoning-with-citations flow, **local and free**, with your data never leaving the machine |\n| **Perplexica self-hosted** | the UX Perplexica has plus a CoVe verifier, FLARE active retrieval, adaptive compute router, and Claude-plugin packaging |\n| **Khoj** | stronger research-specific reasoning (we're not personal-knowledge-focused), six domain presets, and an MCP server for other agents to call |\n| **gpt-researcher** | newer pipeline architecture, better small-model handling, observable trace, plugin ecosystem |\n| **MiroThinker-H1 / OpenResearcher-30B** | they're stronger on BrowseComp; we run on a laptop with no GPU and cost $0 |\n| **Writing your own LangGraph research agent** | save 2-3 months; reuse our 8-node pipeline + 30+ tested env gates + 137 tests |\n\n**Honest read:** on complex multi-hop reasoning benchmarks, Gemma 3 4B\nsits 15–25% below 30 B+ open models. We don't claim to beat GPT-5.4\nPro. We claim to be the best **$0, runs-on-your-laptop, fully-open**\nresearch agent in April 2026.\n\n---\n\n## Quickstart — Mac local\n\n### Option A — PyPI (fastest)\n\n```bash\n# 1) Local inference (Ollama + Gemma 3 4B + embedding model — 3.6 GB combined)\nbrew install ollama\nollama pull gemma3:4b nomic-embed-text\n\n# 2) Self-hosted meta-search (Docker; optional but recommended)\ndocker run -d --name searxng -p 8888:8080 searxng/searxng\n\n# 3) The engine itself\npip install agentic-research-engine\n\n# 4) Go\nexport OPENAI_BASE_URL=http://localhost:11434/v1 OPENAI_API_KEY=ollama\nexport MODEL_SYNTHESIZER=gemma3:4b EMBED_MODEL=nomic-embed-text\nexport SEARXNG_URL=http://localhost:8888\nagentic-research ask \"what is Anthropic's contextual retrieval?\" --domain papers\n```\n\n### Option B — from source\n\n```bash\n# 1) Same local-inference prereqs as Option A (ollama pull + docker run)\n\n# 2) Clone + install (gives you the CLI, TUI, Web GUI, MCP server, benchmarks, tutorials)\ngit clone https://github.com/TheAiSingularity/agentic-research-engine-oss\ncd agentic-research-engine-oss\n(cd scripts/searxng && docker compose up -d)\ncd engine && make install\nmake smoke # end-to-end run on the canonical \"what is contextual retrieval\" question\n```\n\nExpected wall-clock on an M-series Mac: **~45 s** for a factoid,\n~90 s for multi-hop synthesis. Zero dollars per query.\n\n### Higher honesty — cloud-model mode\n\nGemma 3 4B is surprisingly good at **structure** (plan, route, verify,\ncompress) but confabulates **specific factoids** when SearXNG doesn't\nsurface a source containing the right token. Live SimpleQA-mini run on\n2026-04-21 (see [`engine/benchmarks/RESULTS.md`](engine/benchmarks/RESULTS.md))\nshowed `gemma3:4b` emitting \"2023\" for *\"year Anthropic published\nContextual Retrieval\"* (gold: 2024) and \"LayoutLMv3\" for *\"which\ncross-encoder for reranking\"* (gold: bge-reranker-v2-m3).\n\nThe fix you probably want isn't a smarter synthesizer — it's a\n**more honest** one. A 5-question head-to-head on the same retrieval\noutput showed `gpt-5-nano` + `gpt-5-mini` refuse to confabulate when\nevidence was missing (*\"The provided evidence does not answer this\nquestion\"*), where `gemma3:4b` confidently guessed. Per-claim\nfaithfulness went from 82.9 % → 100 %. Pass rate barely moved (1/5\nvs 0/5) because **retrieval is the real bottleneck** — if SearXNG\ndidn't return a source with the gold token, neither model can\nproduce it.\n\nSwap the whole stack to a cloud endpoint:\n\n```bash\n# drop the Ollama base URL (fall back to OpenAI cloud)\nunset OPENAI_BASE_URL\nexport OPENAI_API_KEY=sk-...\n# defaults are already cloud-sized: gpt-5-nano for plan/verify, gpt-5-mini for synth.\n# Explicit override if you want to pin them:\nexport MODEL_PLANNER=gpt-5-nano\nexport MODEL_SYNTHESIZER=gpt-5-mini # or gpt-5, claude-sonnet-4-5, etc.\nagentic-research ask \"…\" --domain papers\n```\n\nCost is dominated by synthesizer tokens (~5–15 k per query). Full\ncloud mode with `gpt-5-nano` + `gpt-5-mini` runs roughly\n**$0.02–0.05 per research query** and is ~2-3× slower than Gemma\nlocal (measured: 127 s vs 52 s mean wall on the 5-question subset).\nWorks with any OpenAI-compatible endpoint — Groq, Together, Mistral,\nDeepSeek, local vLLM — so you can pick a cheap fast model\n(`llama-3.3-70b` on Groq ≈ $0.003/query) or a frontier one. Per-node\nbase-URL routing (run gemma3:4b locally for plan/verify AND gpt-5-mini\non cloud for synth in the same query) is tracked for 0.2; today the\npipeline uses one global `OPENAI_BASE_URL`.\n\n**The bigger accuracy lever is retrieval.** Point\n`LOCAL_CORPUS_PATH` at an indexed corpus containing your answer and\neither model will be correct.\n\n---\n\n## Quickstart — no install (Google Colab)\n\nFive runnable notebooks in [`tutorials/`](tutorials/):\n\n1. [**01 — Engine API quickstart** (mocked, no key)](tutorials/01_engine_api_quickstart.ipynb) — see how the pipeline works without running inference.\n2. [**02 — Groq cloud inference** (free tier)](tutorials/02_groq_cloud_inference.ipynb) — real LLM, no local GPU.\n3. [**03 — Build your own corpus**](tutorials/03_build_your_own_corpus.ipynb) — upload PDFs, index them, query.\n4. [**04 — MCP server from Python**](tutorials/04_mcp_server_from_python.ipynb) — drive the engine as a tool from another agent.\n5. [**05 — Domain presets showcase**](tutorials/05_domain_presets_showcase.ipynb) — compare presets on the same question.\n\nEach notebook is self-contained, runs end-to-end on Colab free tier, no\ncredit card required.\n\n---\n\n## Three ways to drive it\n\n### CLI\n\n```bash\nengine ask \"what is hybrid retrieval?\" --domain papers --memory session\nengine reset-memory\nengine domains list\nengine version\n```\n\n### TUI (Textual — keyboard-driven, SSH-safe)\n\n```bash\nmake tui\n```\n\nThree panes: sources · answer + hallucination flags · trace + memory hits.\nPress Enter to ask, Ctrl-M to cycle memory mode,\nCtrl-L to clear, Ctrl-Q to quit.\n\n### Web GUI (FastAPI + HTMX on `localhost:8080`)\n\n```bash\nmake gui\n# open http://127.0.0.1:8080 in your browser\n```\n\nNo auth. No cloud. No analytics. Dark theme. Streams tokens in place.\n\n---\n\n## What ships\n\n### `engine/` — the flagship\n\n8-node LangGraph pipeline with 2026-SOTA composition:\n`classify → plan → search → retrieve → fetch_url → compress → synthesize → verify`\n\nEvery stage is env-toggleable for leave-one-out ablation. Techniques\nfolded in: HyDE, CoVe verification, iterative retrieval, FLARE active\nretrieval, question classifier router, step critic (ThinkPRM pattern),\nLongLLMLingua-lite compression, cross-encoder rerank\n(`BAAI/bge-reranker-v2-m3`), Anthropic contextual chunking, W6 small-\nmodel hardening (three-case synthesize prompt + per-chunk char cap).\n\n### `core/rag/` — reusable retrieval primitives (v1 stable)\n\n`HybridRetriever` (BM25 + dense + RRF) · `CrossEncoderReranker` ·\n`contextualize_chunks` (Anthropic pattern) · `CorpusIndex` (bring-\nyour-own-PDFs). 5 exports, used by the engine and the archived\nrecipes.\n\n### `archive/recipes/` — pre-engine reference recipes\n\n`research-assistant`, `trading-copilot`, `document-qa`,\n`rust-mcp-search-tool`. All still work; all tests still pass. The\n`research-assistant/production/main.py` is a thin shim over\n`engine.core.pipeline` so the cookbook framing is preserved.\n\n---\n\n## Domain presets\n\nSix YAML files in `engine/domains/`:\n\n| preset | when to use |\n|---|---|\n| `general` | default; anything |\n| `medical` | disease / treatment / drug / trial (PubMed / Cochrane / NEJM bias; no prescriptive advice) |\n| `papers` | academic CS / ML / physics / biology (arXiv + Semantic Scholar + OpenReview) |\n| `financial` | SEC filings, earnings, company fundamentals (dates on every number) |\n| `stock_trading` | technical + news per ticker — **hard rule: never recommends buy/sell/hold** |\n| `personal_docs` | Q&A over your own corpus, air-gapped (only `corpus://` URLs allowed) |\n\nWrite your own in ~10 lines of YAML — see [`docs/domains.md`](docs/domains.md).\n\n---\n\n## Bring your own documents\n\n```bash\npython scripts/index_corpus.py build ~/papers --out ~/papers.idx\nexport LOCAL_CORPUS_PATH=~/papers.idx\nengine ask \"what do my papers say about contextual retrieval?\" --domain personal_docs\n```\n\nSupported formats: PDF (via pypdf), Markdown, plain text, HTML (via\ntrafilatura). The index persists as a directory with a human-readable\n`manifest.json` + a pickled `index.pkl`. Rebuild anytime the docs change.\n\nDetails: [`docs/self-learning.md`](docs/self-learning.md) covers the\ntrajectory + memory model; [`docs/plugins-skills.md`](docs/plugins-skills.md)\ncovers external plugins.\n\n---\n\n## MCP + Claude plugin\n\n`engine/mcp/server.py` is a Python MCP server exposing:\n- `research(question, domain?, memory?)` → structured `{answer, verified_claims, unverified_claims, sources, trace, totals, memory_hits}`\n- `reset_memory()`\n- `memory_count()`\n\nBundled Claude plugin at `engine/mcp/claude_plugin/` — four skills\n(`/research`, `/cite-sources`, `/verify-claim`, `/set-domain`), ready to\nsubmit to the Anthropic marketplace.\n\nRegister in Claude Desktop:\n\n```jsonc\n// ~/Library/Application Support/Claude/claude_desktop_config.json\n{\n \"mcpServers\": {\n \"engine\": {\n \"command\": \"python\",\n \"args\": [\"-m\", \"engine.mcp.server\"],\n \"env\": {\n \"OPENAI_BASE_URL\": \"http://localhost:11434/v1\",\n \"OPENAI_API_KEY\": \"ollama\",\n \"MODEL_SYNTHESIZER\": \"gemma3:4b\",\n \"SEARXNG_URL\": \"http://localhost:8888\"\n }\n }\n }\n}\n```\n\n---\n\n## Plugin / skill loader\n\nInstall third-party Claude plugins or Hermes (`agentskills.io`) skills:\n\n```bash\nengine plugins install gh:owner/some-research-plugin@v1\nengine plugins install file:./my-local-plugin\nengine plugins install https://example.com/marketplace.json\nengine plugins list\nengine plugins uninstall some-plugin\n```\n\nSafety: every install runs a forbidden-symbols scan\n(`eval(`, `exec(`, `os.system(`, …) — rejects plugins that would\nexecute arbitrary code. Registry lives at\n`~/.agentic-research/plugins/`, fully inspectable, wipable.\n\nFull docs: [`docs/plugins-skills.md`](docs/plugins-skills.md).\n\n---\n\n## Architecture at a glance\n\n```\n ┌─────────────┐\n │ question │\n └──────┬──────┘\n ▼\n ┌─────────────────────────┐ T4.3 router — route by question type\n │ classify │\n └──────────┬──────────────┘\n ▼\n ┌─────────────────────────┐ T1 decompose · T2 HyDE · T4.1 critic\n │ plan │ T4.5 refine-on-reject\n └──────────┬──────────────┘\n ▼\n ┌─────────────────────────┐ SearXNG parallel × N\n │ search │ + W5 local corpus (optional)\n │ (+ T4.1 critic) │ + T4.1 coverage critic\n └──────────┬──────────────┘\n ▼\n ┌─────────────────────────┐ T1 hybrid BM25 + dense + RRF\n │ retrieve │ W4.1 cross-encoder rerank (opt-in)\n │ (+ W4.1 rerank) │\n └──────────┬──────────────┘\n ▼\n ┌─────────────────────────┐ W4.2 trafilatura clean-text\n │ fetch_url │ skips corpus:// URLs\n └──────────┬──────────────┘\n ▼\n ┌─────────────────────────┐ T4.4 LLM distillation\n │ compress │ + W6.2 per-chunk char cap\n │ (+ W6.2 cap) │\n └──────────┬──────────────┘\n ▼\n ┌─────────────────────────┐ T2 synth · T4.2 FLARE on hedges\n │ synthesize │ W6.1 three-case anti-hallucinate\n │ (+ FLARE + stream) │ W7 streaming\n └──────────┬──────────────┘\n ▼\n ┌─────────────────────────┐ T2 CoVe — decompose + verify\n │ verify │\n └────────┬────────────────┘\n │\n verified? ── yes ──▶ END\n │\n no\n │\n ◀────── re-search unverified claims ──── loop (bounded by MAX_ITERATIONS)\n```\n\nEvery stage has an `ENABLE_*` flag so you can leave-one-out ablate.\nDeep spec: [`docs/architecture.md`](docs/architecture.md).\n\n---\n\n## Repo layout\n\n```\nagentic-research-engine-oss/\n├── engine/ the flagship research engine\n│ ├── core/ pipeline · models · trace · memory\n│ │ ├── pipeline.py · compaction · domains · plugins\n│ │ ├── models.py\n│ │ ├── trace.py\n│ │ ├── memory.py\n│ │ ├── compaction.py\n│ │ ├── domains.py\n│ │ └── plugins.py\n│ ├── interfaces/\n│ │ ├── cli.py rich stdout CLI with subcommands\n│ │ ├── tui.py Textual TUI\n│ │ └── web/ FastAPI + HTMX localhost GUI\n│ ├── mcp/\n│ │ ├── server.py Python FastMCP server\n│ │ └── claude_plugin/ submittable Claude plugin bundle\n│ ├── domains/ 6 YAML presets\n│ ├── examples/ 5 worked research examples\n│ ├── benchmarks/ mini SimpleQA + BrowseComp fixtures + runner\n│ └── tests/ pytest suite (all mocked, zero-network)\n├── core/rag/ shared retrieval primitives (stable v1)\n├── archive/ pre-engine recipes (kept for reference)\n├── tutorials/ 5 Google Colab notebooks\n│ ├── 01_engine_api_quickstart.ipynb\n│ ├── 02_groq_cloud_inference.ipynb\n│ ├── 03_build_your_own_corpus.ipynb\n│ ├── 04_mcp_server_from_python.ipynb\n│ └── 05_domain_presets_showcase.ipynb\n├── scripts/\n│ ├── searxng/ self-hosted meta-search (docker-compose)\n│ ├── setup-local-mac.sh Ollama + Docker + SearXNG one-liner\n│ ├── setup-vm-gpu.sh Linux + vLLM/SGLang setup\n│ └── index_corpus.py build a CorpusIndex from PDFs/md/txt\n├── docs/\n│ ├── architecture.md deep technical spec\n│ ├── plugins-skills.md write + install plugins\n│ ├── domains.md write a new preset\n│ ├── self-learning.md trajectory logging + memory\n│ ├── progress.md wave-by-wave build log\n│ ├── how-it-works.md elevator pitches + SOTA comparison\n│ ├── launch-checklist.md go-live sequence\n│ └── launch-copy.md drafted HN / Reddit / Twitter copy\n├── .github/\n│ ├── workflows/\n│ │ └── engine-tests.yml CI: mocked suite on every PR\n│ ├── ISSUE_TEMPLATE/\n│ └── PULL_REQUEST_TEMPLATE.md\n├── CONTRIBUTING.md\n├── CHANGELOG.md\n├── CODE_OF_CONDUCT.md\n├── LICENSE MIT\n└── README.md you're reading it\n```\n\n---\n\n## Configuration (env vars)\n\nFull list in `engine/core/pipeline.py` header. Most-common knobs:\n\n| var | default | purpose |\n|---|---|---|\n| `OPENAI_BASE_URL` | unset (cloud OpenAI) | route to Ollama / vLLM / Groq / etc. |\n| `OPENAI_API_KEY` | `ollama` | sentinel for local; real key for cloud |\n| `MODEL_SYNTHESIZER` | `gpt-5-mini` (cloud) or `gemma3:4b` (Mac-local path) | final-answer model. Swap to `gpt-5`, `claude-sonnet-4-5`, `llama-3.3-70b` on Groq, etc., for higher factoid accuracy while keeping the rest of the pipeline local. |\n| `TOP_K_EVIDENCE` | auto (5 for small, 8 for large models) | retrieval budget |\n| `ENABLE_RERANK` | `0` | opt-in; first run downloads bge-reranker-v2-m3 (~560 MB) |\n| `ENABLE_FETCH` | `1` | trafilatura full-page fetch |\n| `ENABLE_STREAM` | `1` | stream synthesis tokens to stdout |\n| `ENABLE_TRACE` | `1` | per-call observability + summary at CLI end |\n| `LOCAL_CORPUS_PATH` | unset | set to an index dir to augment search with your docs |\n| `MEMORY_DB_PATH` | `~/.agentic-research/memory.db` | SQLite trajectory store |\n\nFull list: [`docs/architecture.md`](docs/architecture.md) env-vars section.\n\n---\n\n## Testing\n\n```bash\ncd engine && make test # 120+ mocked tests in engine/tests/\n# or repo-wide:\nPYTHONPATH=$(pwd) .venv/bin/python -m pytest core/rag recipes engine/tests -q\n```\n\nAll tests are mocked — no network, no API key, no model downloads. Live\nintegration smokes are separate (`make smoke`).\n\nCI runs on every push / PR touching engine / core / recipes — see\n[`.github/workflows/engine-tests.yml`](.github/workflows/engine-tests.yml).\n\n---\n\n## Troubleshooting\n\n| symptom | likely cause | fix |\n|---|---|---|\n| `ModuleNotFoundError: No module named 'engine'` | `PYTHONPATH` missing the repo root | `export PYTHONPATH=$(pwd)` from the repo root |\n| CLI answer is empty + fast | Ollama not running | `ollama serve` in another terminal, or `ollama list` to check |\n| `Connection refused on :8888` | SearXNG not up | `cd scripts/searxng && docker compose up -d` |\n| `Connection refused on :11434` | Ollama not running | `ollama serve`, or let the system service start it |\n| First `make smoke` hangs ~20 s before output | Model warming up on first request | normal; subsequent queries are faster |\n| `ENABLE_RERANK=1` stalls on first run | 560 MB bge-reranker download | wait it out once; cached after |\n| `[corpus] LOAD BROKEN` | corrupt or wrong-version index | delete + rebuild via `scripts/index_corpus.py` |\n| TUI shows gibberish over SSH | terminal too narrow | resize to ≥ 100 cols; Textual needs space for the 3-pane layout |\n| Web GUI shows `Invalid memory mode` | malformed POST | use the form UI; values validated against `off/session/persistent` |\n| Streaming cuts off mid-answer | flaky backend | re-run; batched fallback kicks in on next attempt. Set `ENABLE_STREAM=0` if it persists |\n| `zsh: command not found: twine` (or similar) after `uv pip install ` | uv's venv isn't auto-activated by your shell | use `.venv/bin/ …`, `uv run …`, or `source .venv/bin/activate` before running |\n| `bad interpreter: .../python3: no such file or directory` after moving or renaming the repo dir | venv shebangs are absolute paths tied to the dir the venv was created in | recreate: `rm -rf .venv && uv venv && uv pip install -e .` (or re-install whatever you had) |\n| `make test` says 0 tests collected | wrong CWD | run from the `engine/` dir or set `PYTHONPATH` |\n| Claude Desktop doesn't see the plugin | plugin.json in wrong path | `/plugin marketplace add /engine/mcp/claude_plugin` |\n\nStill stuck? Open an issue with the [`bug_report`](.github/ISSUE_TEMPLATE/bug_report.md)\ntemplate — include `ollama list`, `engine version`, and the error.\n\n---\n\n## Honest limits\n\n- **Gemma 4B ≠ GPT-5.4 Pro.** 15–25 % below 30 B+ open models on hard\n multi-hop. We position as \"best $0 local\", not \"SOTA.\"\n- **Gemma 3 4B confabulates specific factoids** when SearXNG doesn't\n return a source that contains the right token. Measured on\n SimpleQA-mini: 0/20 strict pass rate (see\n [`engine/benchmarks/RESULTS.md`](engine/benchmarks/RESULTS.md) —\n `verified_ratio` 85.5 %, zero `must_not_contain` hits; the model\n isn't emitting *banned* strings, it's picking wrong ones). Mitigations:\n (a) swap the whole stack to a cloud endpoint (see \"Higher factoid\n accuracy\" above — `$0.02–0.05/query` with `gpt-5-nano` + `gpt-5-mini`),\n (b) give the engine a `LOCAL_CORPUS_PATH` so your own docs become\n retrieval targets, (c) set `ENABLE_RERANK=1` to bias retrieval\n toward the right sources.\n- **CoVe confirms internal consistency, not ground truth.** Every\n synthesized claim is checked against retrieved evidence; claims\n don't get verified *by the world*. If retrieval misses, CoVe will\n still happily verify a confidently-wrong answer. The engine will\n never fabricate citations, but it can confidently repeat wrong\n information that was in its evidence pool.\n- **No LoRA fine-tuning in v1.** Trajectory data is collected; actual\n model training deferred until GPU access + data volume.\n- **No hosted SaaS.** Local-first is the entire v1 positioning.\n- **Team / multi-user features.** Out of scope for v1.\n- **General web crawler / own search index.** Not shipping. SearXNG\n stays. A curated research-focused index may land in v2.\n- **Mobile.** Not in scope.\n\n---\n\n## Status + roadmap\n\n- **0.1.3 — public alpha** (current). Features listed above; on PyPI +\n the official MCP registry + the Anthropic plugin marketplace. See\n [`CHANGELOG.md`](CHANGELOG.md).\n- **0.2** — specialist tool wiring (`tools_enabled` field in presets finally activates), first LoRA run if GPU arrives, plugin catalog in `docs/`.\n- **0.3** — team-collab features (shared memory, PR-driven domain presets), desktop app packaging via Tauri.\n- **0.4+** — open-work tracked in [GitHub Issues](https://github.com/TheAiSingularity/agentic-research-engine-oss/issues).\n\n---\n\n## Contributing\n\nGood first issues: [`CONTRIBUTING.md`](CONTRIBUTING.md). RFCs for\nanything pipeline-scope. Plugin + domain-preset submissions welcome.\n\nNo Co-Authored-By trailers; author-as-written-by.\n\n---\n\n## License\n\nMIT. See [`LICENSE`](LICENSE).\n\n### Related (sibling projects)\n\n- [HermesClaw](https://github.com/TheAiSingularity/hermesclaw) — the secure runtime these recipes can run inside\n- [NVIDIA/OpenShell](https://github.com/NVIDIA/OpenShell) — kernel-level agent sandbox\n- [NousResearch/hermes-agent](https://github.com/NousResearch/hermes-agent) — self-improving agent (whose `agentskills.io` skill format we interoperate with)\n\n---\n\n### MCP registry ownership\n\nThis PyPI package is the official source of the MCP server registered at\n. The line below is the\nownership marker the registry validates — **do not remove** when\nediting this README.\n\nmcp-name: io.github.TheAiSingularity/agentic-research\n", "bytes": 25105, "sha": "f85bc7254357f3e349febf73f9dc2e9e97c6c62ec1343a742e6aeeaccaa3df3d", "repo_slug": "theaisingularity/agentic-research-engine-oss", "fonte": "repo", "truncated": false, "api": "https://agentalog.com/api/listings/mcp_io_github_theaisingularity_agentic_resea_0c31220a/readme" }