ScholarAgent

A multi-agent scientific literature research system that discovers papers on arXiv, Semantic Scholar, and GitHub, analyzes them through a 5-agent pipeline, and persists the findings in a semantic memory layer exposed to your coding agent as an MCP server.

The 30-second version

Ask your editor "what's the state of the art on RLHF reward models?" and ScholarAgent will:

1. Search arXiv + Semantic Scholar, follow citation graphs, pull code examples from GitHub.

2. Have specialist agents (Reader, Critic, Analyst, Synthesizer) extract claims, score methodology, find themes and contradictions, then write a Markdown literature review.

3. Index every finding with embeddings in a local SQLite database at ~/.scholaragent/memory.db.

4. Expose 9 MCP tools so the next time you ask anything semantically adjacent, memory_lookup returns it instantly — no re-research needed.

A research run goes from a natural-language query to a cited review in ~5 seconds (quick depth) to ~5 minutes (deep depth).

Related MCP server: OpenArx

Architecture at a glance

Your query
   │
   ▼
 Dispatcher (orchestrator, writes Python that calls sub-agents)
   │
   ├─► Scout        search arXiv + Semantic Scholar, follow citations
   ├─► Reader       extract key_claims, methodology, results, limitations
   ├─► Critic       score rigor + relevance, flag bias, rate reliability
   ├─► Analyst      find themes (3+ papers), contradictions, gaps
   └─► Synthesizer  write structured Markdown review with citations
         │
         ▼
 ContextStream (structured pipeline state + per-agent conversation traces)
         │
         ▼
 MemoryStore (SQLite + embeddings, source_types: paper | code | docs | synthesized_report)
         │
         ▼
 MCP Server (9 tools over stdio, registered in Claude Code, Cursor, Windsurf, VS Code)

Each agent runs an RLM loop: the LLM generates Python code in a sandboxed REPL, the code executes, the output feeds back to the LLM, repeat until FINAL() terminates.

Multi-LLM routing. Scout uses a cheap fast model (gpt-4o-mini by default). Reader/Critic/Analyst/Synthesizer use a strong model (claude-sonnet-4-6 by default). Supported backends: OpenAI, Anthropic, and LM Studio (any OpenAI-compatible local server — zero API cost).

The 9 MCP tools

Depth levels:

• quick — Source search + dedup + index. No agents. ~5–10s.

• normal — Scout + Reader + Critic per paper (up to 3 parallel workers). ~30–60s.

• deep — Full Dispatcher orchestrating all 5 agents to produce a synthesized report. ~2–5min.

Focus modes (passed to every agent's system prompt):

• implementation — emphasize code, APIs, how-to.

• theory — emphasize concepts, algorithms, math.

• comparison — emphasize alternatives, benchmarks, trade-offs.

Fallback is visible. Every memory_research response includes both requested_depth and actual_depth. If deep fell back to normal because the dispatcher timed out, you'll see fallback_reason in the response — no silent downgrades.

Installation

Prerequisites

• Python 3.12+

• One of:

  • OPENAI_API_KEY (for embeddings — required by default) plus ANTHROPIC_API_KEY (for the strong model in normal/deep research), or

  • A local LM Studio server (everything free, no API keys needed, see below)

• Optional: GITHUB_TOKEN for GitHub code search (otherwise you'll get 401s — non-fatal).

Install from PyPI (coming soon — currently install from source)

pip install scholaragent
scholaragent-install

The installer auto-detects:

• Claude Code — ~/.claude/settings.json

• Cursor — ~/.cursor/mcp.json

• Windsurf — ~/.windsurf/mcp.json

• VS Code — ~/.vscode/mcp.json

• LM Studio — ~/.lmstudio/mcp.json (same JSON shape as above). Also detects a running LM Studio at http://localhost:1234 and suggests --backend lmstudio.

• Codex CLI — ~/.codex/config.toml (TOML; upserts [mcp_servers.scholar-memory] without touching other sections).

• Docker Desktop MCP Toolkit — if ~/.docker/mcp/ exists, the installer prints the exact docker mcp server add … command to run (Docker manages its MCP registry via CLI, not a static file).

Each JSON-based target gets an mcpServers.scholar-memory entry in its JSON config, while Codex CLI gets a [mcp_servers.scholar-memory] table in ~/.codex/config.toml; both point at the scholaragent-server binary.

Install from source (bash installer)

git clone https://github.com/byBasiliosP/RLM-Agent.git
cd RLM-Agent

# cloud backend — export your keys first
export OPENAI_API_KEY="sk-..."
export ANTHROPIC_API_KEY="sk-ant-..."   # required for normal/deep depth
export GITHUB_TOKEN="ghp_..."           # optional
./install.sh

# or, fully local via LM Studio — no keys required
./install.sh --backend lmstudio

install.sh will:

1. Check Python ≥ 3.12.

2. Create ./.venv/ and pip install -e . into it.

3. Verify the scholaragent-server entry point exists.

4. Choose a backend:

   • --backend lmstudio skips the cloud-key check entirely.

   • If cloud keys are missing and a local LM Studio is running on localhost:1234, the installer offers to switch ([Y/n] prompt). Pass --yes to auto-accept.

5. Delegate registration to scholaragent-install, which upserts the MCP entry in every detected target (Claude Code / Cursor / Windsurf / VS Code / LM Studio / Codex CLI) and prints the exact docker mcp server add … line if Docker Desktop MCP Toolkit is present.

Install from source (manual)

git clone https://github.com/byBasiliosP/RLM-Agent.git
cd RLM-Agent
python3.12 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
scholaragent-install

Install for LM Studio (no API keys)

scholaragent-install --backend lmstudio \
  --strong-model qwen3-30b-a3b \
  --cheap-model llama-3.2-3b-instruct

Then start LM Studio at http://localhost:1234/v1 with both models loaded. To also run embeddings locally, set these in your shell profile:

export SCHOLAR_EMBEDDING_BACKEND=lmstudio
export SCHOLAR_EMBEDDING_MODEL=text-embedding-nomic-embed-text-v1.5

Uninstall

scholaragent-install --uninstall
# or
./install.sh --uninstall

Removes the scholar-memory entry from every agent config it finds. The ~/.scholaragent/memory.db stays put — delete it manually if you want a clean slate.

Restart your coding agent

Every coding agent caches MCP configuration at startup. After install, fully restart (not just reload window) so the 9 new tools appear.

Configuration

All config is environment variables — no config files.

Python API

Skip MCP and use directly:

from scholaragent import ScholarAgent

agent = ScholarAgent(
    strong_model={"backend": "anthropic", "model_name": "claude-sonnet-4-6"},
    cheap_model={"backend": "openai", "model_name": "gpt-4o-mini"},
    max_iterations=15,
    verbose=True,
)

result = agent.research("State of the art on RLHF reward models")
print(result.result)   # Markdown literature review

For a lower-level entrypoint that matches what the MCP server does internally:

from scholaragent.runtime import RuntimeContainer
from pathlib import Path

container = RuntimeContainer(
    data_dir=Path.home() / ".scholaragent",
    db_path=str(Path.home() / ".scholaragent/memory.db"),
    model_config={
        "strong": {"backend": "anthropic", "model_name": "claude-sonnet-4-6"},
        "cheap":  {"backend": "openai",    "model_name": "gpt-4o-mini"},
    },
)
pipeline = container.get_pipeline()
container.ensure_pipeline_agents()
result = pipeline.run("RLHF reward models", depth="deep", focus="implementation")
print(result["actual_depth"], result["entries_added"])
container.close()

Things you should know

Honest notes about warts, limits, and design choices.

Costs

• Quick depth is cheap — just embedding costs. A single research run embeds maybe 15 entries, which is a fraction of a cent on text-embedding-3-small.

• Normal depth runs Reader + Critic per paper, up to 3 in parallel. Expect ~$0.05–0.30 per run on default models.

• Deep depth runs the full pipeline with iteration loops. Expect ~$0.50–3.00 per run on default models.

• Set an API spending limit on your OpenAI/Anthropic accounts. The dispatcher has token budgets but they're per-run, not per-day.

The Linter agent exists but isn't wired up

scholaragent/agents/linter.py implements LinterAgent (static analysis over a code path), and there are tests for it, but it is not registered in the default AgentRegistry in either scholaragent/__init__.py or scholaragent/runtime.py. Calling call_agent("linter", ...) from the Dispatcher will raise because the registry has no entry for it. If you want linter integration, register it yourself or open a PR.

Source adapter caveats

• GitHub code search is hardcoded to language="python" in the default SourceCollector. You can change it via SourceCollector(default_code_language="rust") or pass code_language= to collect(), but the MCP tool path uses the default.

• Docs search (sources/docs.py::search_docs) currently constructs a single URL: https://docs.python.org/3/search.html?q={query}. It is a stub. fetch_docs(url) against arbitrary URLs works fine — the search part doesn't.

• Semantic Scholar rate-limits aggressively (429). Errors are caught and returned in the errors list; results from other sources still come through.

• arXiv redirects HTTP→HTTPS; httpx follows redirects transparently but adds a small latency on the first call.

Memory store scalability

• Vector search is cosine similarity computed in Python against all stored embeddings. Fine for thousands of entries. If you index hundreds of thousands, consider swapping MemoryStore.search for a real vector index (FAISS, sqlite-vss, etc.) — the EmbeddingBackend is an ABC to make this easier.

• The SQLite file uses WAL mode. Multiple readers are fine; heavy concurrent writes will serialize through a single lock.

Dedup behavior

• memory_research caches results for 7 days by default. Calling the same query twice inside that window returns {"status": "cached"} instead of re-running. Pass force=True to override.

• Within a run, papers are deduped by arxiv_id first, then normalized title. S2 entries win over arXiv entries on collision because they carry citation counts.

Research pipeline fallback

• Deep falls back to normal if the Dispatcher errors or returns no result.

• Normal falls back to quick if Scout fails.

• Quick is the floor — it never falls back.

• The returned dict carries requested_depth, actual_depth, and fallback_reason so the caller sees exactly what happened. The legacy depth key aliases actual_depth for backwards compatibility.

MCP transport

• scholaragent-server speaks MCP over stdio — it's spawned by the coding agent as a subprocess. There is no HTTP endpoint.

• The 9 tools are defined in scholaragent/_manifest.py::MCP_TOOLS (single source of truth). A drift-guard test parses the @mcp.tool() decorators and fails the build if the manifest diverges.

Testing

• 609 tests across 36 files.

• pytest-asyncio is installed (transitive dep) but disabled via pyproject.toml: addopts = "-p no:asyncio". There are no async tests, and the plugin deadlocks with threading.Lock-based lazy init.

• Run: python -m pytest tests/ -v

• tests/test_installer.py assumes a repo-level .venv/ and fails in worktree checkouts. tests/test_web_tools_live.py hits real search engines and is skipped by default.

Security

• API keys are read from the shell environment at runtime. They are never written to any config file the installer touches.

• The REPL has restricted builtins (no input, eval, exec, compile). RESERVED_NAMES (FINAL, FINAL_VAR, call_agent, llm_query, etc.) are restored after every code execution so the LLM's generated code can't corrupt the scaffold.

Project structure

scholaragent/
├── __init__.py                  # Public API: ScholarAgent class
├── _manifest.py                 # MCP_TOOLS tuple — single source of truth
├── mcp_server.py                # FastMCP server (9 tools, stdio)
├── runtime.py                   # RuntimeContainer — owns store/pipeline/agent-infra lifecycle
├── installer.py                 # CLI: scholaragent-install (reads _manifest.MCP_TOOLS)
├── core/                        # SpecialistAgent ABC, Dispatcher, registry, LMHandler, ContextStream
├── agents/                      # scout, reader, critic, analyst, synthesizer, linter (not registered)
├── clients/                     # OpenAI, Anthropic, LM Studio, router, token counter, rate limiter
├── environments/                # Sandboxed LocalREPL
├── memory/
│   ├── store.py                 # SQLite + cosine search
│   ├── types.py                 # MemoryEntry, ResearchLogEntry, VALID_SOURCE_TYPES
│   ├── embeddings.py            # EmbeddingBackend ABC + OpenAIEmbeddings + LRU cache
│   ├── research.py              # ResearchPipeline — depth orchestration
│   ├── source_collector.py      # SourceCollector — raw retrieval (arXiv, S2, GitHub, docs)
│   ├── indexer.py               # ResultIndexer — MemoryEntry construction + store writes
│   └── research_result.py       # ResearchResult dataclass (requested_depth, actual_depth)
├── tools/                       # arxiv, semantic_scholar, web, pdf_extractor, quality
├── sources/                     # github code search, html doc fetcher
└── utils/                       # parsing, prompts, retry, budget, cost, llm cache
tests/                           # 609 tests
install.sh                       # Bash installer (wraps scholaragent-install)

Development

pip install -e ".[dev]"
python -m pytest tests/ -v                                  # full suite
python -m pytest tests/test_research_pipeline.py -v         # one file
python -m pytest tests/ --ignore=tests/test_installer.py \
                       --ignore=tests/test_web_tools_live.py  # skip env-specific

Citations & Acknowledgements

Research Foundations

ScholarAgent is built on ideas from the following academic work:

• Recursive Language Models (RLM) — Zhang, A. L., Kraska, T., & Khattab, O. (2026). Recursive Language Models. arXiv:2512.24601. https://arxiv.org/abs/2512.24601 The core REPL-driven orchestration paradigm. Every agent loop in ScholarAgent follows the RLM pattern of code generation → execution → observation → iteration.

• ReAct — Yao, S., Zhao, J., Yu, D., et al. (2023). ReAct: Synergizing Reasoning and Acting in Language Models. ICLR 2023. arXiv:2210.03629. https://arxiv.org/abs/2210.03629 The thought–action–observation loop that underpins agent reasoning.

• Multi-Agent Systems — Guo, T., et al. (2024). Large Language Model based Multi-Agents: A Survey of Progress and Challenges. IJCAI 2024. arXiv:2402.01680. https://arxiv.org/abs/2402.01680; Li, S., et al. (2024). arXiv:2412.17481. https://arxiv.org/abs/2412.17481 Surveys informing the five-agent specialist architecture.

Frameworks & Protocols

• RLM — REPL-driven LM orchestration patterns (Zhang et al., MIT CSAIL)

• MCP — Model Context Protocol for agent interoperability (Anthropic, 2024)

• FastMCP — Python MCP server framework by Jeremiah Lowin

Data Sources

• arXiv API — Open access to scientific papers (operated by Cornell University)

• Semantic Scholar Academic Graph API — Paper metadata and citation graphs (Allen Institute for AI)

• GitHub REST API — Code search (GitHub / Microsoft)

Key Libraries

• OpenAI Python SDK — LLM and embedding API client (OpenAI)

• Anthropic Python SDK — LLM API client (Anthropic)

• httpx — Async HTTP client (Encode)

• Rich — Terminal formatting (Will McGugan / Textualize)

• NumPy — Numerical operations (NumPy community)

• pypdf — PDF text extraction (pypdf community)

• Trafilatura — Web text extraction (Adrien Barbaresi)

License

Citations & Acknowledgements

Research Foundations

ScholarAgent is built on ideas from the following academic work:

• Recursive Language Models (RLM) — Zhang, A. L., Kraska, T., & Khattab, O. (2026). Recursive Language Models. arXiv:2512.24601. https://arxiv.org/abs/2512.24601 The core REPL-driven orchestration paradigm. Every agent loop in ScholarAgent follows the RLM pattern of code generation → execution → observation → iteration.

• ReAct — Yao, S., Zhao, J., Yu, D., et al. (2023). ReAct: Synergizing Reasoning and Acting in Language Models. ICLR 2023. arXiv:2210.03629. https://arxiv.org/abs/2210.03629 The thought–action–observation loop that underpins agent reasoning.

• Multi-Agent Systems — Guo, T., et al. (2024). Large Language Model based Multi-Agents: A Survey of Progress and Challenges. IJCAI 2024. arXiv:2402.01680. https://arxiv.org/abs/2402.01680; Li, S., et al. (2024). arXiv:2412.17481. https://arxiv.org/abs/2412.17481 Surveys informing the five-agent specialist architecture.

Frameworks & Protocols

• RLM — REPL-driven LM orchestration patterns (Zhang et al., MIT CSAIL)

• MCP — Model Context Protocol for agent interoperability (Anthropic, 2024)

• FastMCP — Python MCP server framework by Jeremiah Lowin

Data Sources

• arXiv API — Open access to scientific papers (operated by Cornell University)

• Semantic Scholar Academic Graph API — Paper metadata and citation graphs (Allen Institute for AI)

• GitHub REST API — Code search (GitHub / Microsoft)

Key Libraries

• OpenAI Python SDK — LLM and embedding API client (OpenAI)

• Anthropic Python SDK — LLM API client (Anthropic)

• httpx — Async HTTP client (Encode)

• Rich — Terminal formatting (Will McGugan / Textualize)

• NumPy — Numerical operations (NumPy community)

• pypdf — PDF text extraction (pypdf community)

• Trafilatura — Web text extraction (Adrien Barbaresi)

License

MIT