Quantum Research Hub Server

A local-first, MCP-powered autonomous research system for quantum computing. It watches arXiv, builds compact paper memory, proposes small testable ideas, generates and runs bounded experiments, validates the results skeptically, and shows everything in a dashboard — all under a budget, with every agent action logged.

Every day it answers: what changed in quantum computing, which papers matter, what ideas are worth testing, and what the agents actually tried.

This is a functional MVP, not a skeleton: arXiv search is real (with a clean injectable adapter for tests), the experiment engine runs a genuine tiny VQE against exact diagonalization, and the validator can reject results.

---

Highlights

• Researcher MCP server (FastMCP) exposing 17 bounded tools to Claude Code.
• Real arXiv ingestion (httpx + feedparser) -> compact paper cards with deterministic fallback and an optional Claude model pass.
• SQLite storage for papers, chunks, ideas, experiments, runs, agent events, and budget events.
• Local vector paper-memory search using dependency-free hashed embeddings (QRH_MEMORY_BACKEND=bm25|hybrid remains available).
• Orchestrator + 8 agents: Paper Scout, Summarizer, Curator, Idea Generator, Experiment Builder, Runner, Validator/Critic, Reporter.
• Daily run under low/medium/high budget profiles -> daily report; scheduler can also write weekly reports.
• Real experiment engine: tiny TFIM VQE and tensor-network-structured ansatz templates (numpy), run in a sandboxed, timeout-bounded, approval-gated subprocess.
• Dashboard: Overview, Papers, Ideas, Experiments, Agent Logs, Budget, Reports — available as a zero-dependency stdlib HTTP app (recommended, always runs) and as a Streamlit app.
• Safety first: small CPU smoke runs are autonomous; installs/GPU/long jobs require approval; secrets are scrubbed from experiment subprocesses; ideas must cite source papers; no experiment is valid without a baseline.

---

Quickstart

The whole MVP runs on a lightweight stack (mcp, pydantic, httpx, feedparser, numpy, pandas, streamlit, pyyaml). pypdf is optional (full-text parsing only) — the MVP works without it.

Option A — uv (recommended)

uv sync
cp .env.example .env
uv run python -m researcher_mcp.storage.db init
uv run python scripts/seed_demo.py            # demo data so the dashboard is populated
uv run python -m orchestrator.daily_run --profile low
uv run python -m apps.dashboard.server        # dashboard → http://127.0.0.1:8533
uv run pytest

Option B — pip / existing interpreter

python -m pip install -e ".[dev]"
cp .env.example .env
python -m researcher_mcp.storage.db init
python scripts/seed_demo.py
python -m orchestrator.daily_run --profile low
python -m apps.dashboard.server               # dashboard → http://127.0.0.1:8533
python -m pytest

Option C — one shot

scripts/bootstrap.sh        # installs deps, makes .env, inits + seeds the DB
scripts/run_daily.sh --profile low
scripts/dev.sh              # seed + launch the dashboard

The shell scripts default to python3; override with PYTHON="uv run python" scripts/run_daily.sh.

---

How to run each piece

Action	Command
Initialize the DB	python -m researcher_mcp.storage.db init
DB status / reset	python -m researcher_mcp.storage.db status · … reset --yes
Seed demo data	python scripts/seed_demo.py
Daily research run	python -m orchestrator.daily_run --profile low
Daily run (full pipeline w/ experiment)	python -m orchestrator.daily_run --profile medium
Daily run (no network/demo)	python -m orchestrator.daily_run --profile low --offline
Weekly report	python -m orchestrator.scheduler weekly --profile low
Scheduler loop	python -m orchestrator.scheduler loop --profile low --weekly
Dashboard (recommended, zero deps)	python -m apps.dashboard.server → http://127.0.0.1:8533
Dashboard (Streamlit)	streamlit run apps/dashboard/Home.py → http://localhost:8501
MCP server	python -m researcher_mcp.server (stdio)
Build the concept map	python -m researcher_mcp.indexing.pipeline build --reset
Concept-map stats	python -m researcher_mcp.indexing.pipeline stats
Export reasoning dataset	python -m researcher_mcp.indexing.dataset_export --format all
Tests	python -m pytest

Dashboard note: the stdlib dashboard (python -m apps.dashboard.server) has zero third-party dependencies and always runs. The Streamlit dashboard is equivalent but imports pyarrow; if your Python env has a pyarrow built for a different NumPy major version (a common conda/pip mismatch), Streamlit will fail to import. A clean uv sync avoids this, or use the stdlib dashboard.

Budget profiles

profile	papers/day	ideas/day	experiments created	experiments run
low	5	3	0	0
medium	15	8	1	1
high	30	15	2	2

On low, the pipeline discovers/ingests/ranks/ideates but does not create or run experiments (cap 0) — by design. Use --profile medium to exercise the full build → run → validate flow. The seed script uses medium.

---

Synthesis engine: conceptual maps → reasoning dataset

Beyond per-paper cards, the hub maintains a concept-map index over the whole corpus and can export it as a first-principles training dataset for scientific reasoning models. Every ingested paper is indexed automatically; you can also (re)build the whole map from stored papers at any time.

Pipeline: papers → concept extraction + term mining → evidence-backed graph → JSONL dataset

• Concepts come from two sources: a curated quantum-computing seed ontology (~95 concepts: methods, ansätze, models, math objects, benchmarks, fields, hardware) and a domain-general term miner (acronym+definition detection, scientific head-noun phrases) that grows the vocabulary from any corpus. Mined candidates are auto-promoted to concepts when well-evidenced.
• Relations are typed (improves_on, generalizes, applies_to, requires, enables, combines, compared_to, benchmarked_on, plus low-confidence co_occurs). Each relation persists the exact evidence sentence + char offsets + a confidence score — the supervision signal for training.
• Storage is a batched SQLite writer (one connection per build, not one per row) plus FTS5 full-text indexes. concept_edges is re-derived from evidence, so re-indexing is idempotent.

python -m researcher_mcp.indexing.pipeline build --reset   # build map from all stored papers
python -m researcher_mcp.indexing.pipeline stats           # concept/edge/evidence counts
python -m researcher_mcp.indexing.dataset_export --format all --min-confidence 0.5

Four exported dataset formats (JSONL in data/datasets/, each record carries provenance back to arXiv IDs + evidence):

format	unit	teaches
triples	(source, relation, target) + evidence sentence	atomic grounded facts
chains	multi-hop reasoning trace, evidence per hop	transitive/compositional reasoning
qa	grounded question/answer over a concept neighborhood	retrieval-grounded answering
contrastive	two papers characterizing the same pair differently	evidence-weighing / disagreement

MCP tools: build_concept_index, get_index_stats, export_reasoning_dataset, plus search_concepts, get_concept_graph, get_bridge_concepts, get_concept_neighborhood, get_top_concepts.

Dataset yield depends on input richness. The relation extractor is deterministic and high-precision but low-recall: on abstracts alone, most concept pairs co-occur without an explicit relation verb in the same sentence, so they become low-confidence co_occurs (filtered at --min-confidence 0.5). Ingesting full-text PDFs (ingest_paper(..., download_pdf=True) with the pdf extra) multiplies typed-relation density. The relation direction is also a nearest-concept heuristic (no dependency parsing) — every row keeps its evidence sentence so direction is correctable downstream or by a later model pass.

---

Using the MCP server from Claude Code

Add the server to Claude Code (stdio). Example .mcp.json / client config:

{
  "mcpServers": {
    "quantum-research-hub": {
      "command": "python",
      "args": ["-m", "researcher_mcp.server"],
      "cwd": "/absolute/path/to/this/repo"
    }
  }
}

Tools exposed: search_arxiv, ingest_paper, get_paper_card, search_paper_memory, list_recent_papers, create_idea, list_ideas, create_experiment_from_idea, get_experiment, list_experiments, run_experiment, get_experiment_results, validate_experiment, create_daily_report, create_weekly_report, get_budget_status, get_overview.

Subagent definitions live in .claude/agents/ (paper-scout, curator, idea-generator, experiment-builder, validator, reporter, architect, test-engineer, plus paper-summarizer, experiment-runner, dashboard-builder, mcp-server-engineer).

---

Safety & approval model

Autonomous: arXiv search, paper-card creation, ranking, idea generation, experiment-folder creation, unit/smoke tests, short CPU runs, dashboard/DB updates.

Requires approval (returns needs_approval, does nothing): package installs, GPU, jobs > the configured timeout / medium+ runner modes, cloud/paid APIs, deleting files outside data/ and experiments/runs/, changing safety logic.

Hard rules: every idea cites ≥1 source paper; every experiment has a baseline, metric, seed, and validator note; the variational energy can never drop below the exact ground state (flagged as a bug); experiment subprocesses get a secret-scrubbed environment and a hard wall-clock timeout.

Set QRH_APPROVAL_GRANTED=1 (or pass approve=True to run_experiment) to allow a single non-small run when you have reviewed it.

---

Project layout

researcher_mcp/        # MCP server + tools + ingestion + storage (the capability layer)
  server.py            # FastMCP server (python -m researcher_mcp.server)
  config.py            # paths, budget profiles, categories, keyword groups
  tools/               # arxiv, paper, memory, idea, experiment, runner, budget, dashboard, concept, dataset
  ingest/              # arxiv_client, paper_card, chunker, concept_extractor, pdf_* (optional)
  indexing/            # writer (batched), term_miner, graph_store, pipeline, dataset_export
  storage/             # schema.sql, db.py, models.py (pydantic), repository.py, concept_graph.py, vector_store.py
orchestrator/          # the autonomous daily layer
  daily_run.py         # python -m orchestrator.daily_run --profile {low,medium,high}
  budget_manager.py    # enforces caps, records usage
  agent_router.py      # RunContext + pipeline
  reporting.py         # daily + weekly report builders
  agents/              # the 8 pipeline agents
apps/dashboard/        # server.py (stdlib HTTP, recommended) + Streamlit Home.py + pages/1..6
experiments/
  templates/vqe_baseline/          # runnable TFIM-VQE template
  templates/tensor_network_ansatz/ # runnable matched-parameter ansatz template
  runs/                            # generated experiment folders
scripts/               # bootstrap, dev, run_daily, reset_dev_db, run_tests, seed_demo
tests/                 # pytest (mocked network)
data/, db/             # artifacts + SQLite (gitignored)
.claude/agents/        # subagent definitions   .claude/settings.json

Layout note: researcher_mcp/ and orchestrator/ are top-level packages (a flattened version of ARCHITECTURE.md's services/... tree) so every documented python -m ... command works from the repo root with no install. Module paths are identical to the architecture doc.

---

What an experiment contains

Each experiments/runs/<id>/ has: experiment.yaml, hypothesis.md, related_papers.json, plan.md, src/run.py, tests/test_smoke.py, configs/config.json, results/metrics.json (+ logs/, plots/), report.md, validator_notes.md. Templates report exact_energy, vqe_energy, baseline_energy, energy_error, improvement_over_baseline, parameter_count, seed_stability_std, and runtime_seconds. The tensor template also reports structured_ansatz_energy, hardware_efficient_energy, and structured_vs_hardware_delta.

---

Configuration (.env)

See .env.example. Common knobs: QRH_DB_PATH, QRH_DATA_DIR, QRH_BUDGET_PROFILE, QRH_LOOKBACK_DAYS, QRH_ARXIV_MIN_INTERVAL, QRH_EXPERIMENT_TIMEOUT_SECONDS, QRH_APPROVAL_GRANTED, and QRH_MEMORY_BACKEND.

Optional Claude model pass:

export QRH_ENABLE_MODEL_PASS=1
export ANTHROPIC_API_KEY=...
export QRH_CLAUDE_MODEL=claude-sonnet-4-5

When disabled or unavailable, paper cards, ideas, and reports fall back to the deterministic path. No secrets are required for the offline MVP.

---

Limitations (MVP)

• The Claude model pass is optional and requires an Anthropic key.
• Circuit-cutting and QML templates are still stubs that fall back to the VQE template.
• The vector memory backend is a local hashed embedding index, not a persistent Chroma/FAISS store yet.
• Full-text PDF parsing requires the optional pdf extra.

See ARCHITECTURE.md for the full design and CLAUDE.md for the working rules.