bbn-mcp-server
Exposes Big Bang Nucleosynthesis computation and analysis tools to LLM agents, enabling them to scan parameters like baryon density, extra relativistic species, and neutron lifetime, and to fit abundances against observations.
README
bbn-mcp-server
An MCP server that exposes science tools — Big Bang Nucleosynthesis, the
particle physics of the universe's first three minutes — to any LLM agent.
Built on the same pattern as
spectra-mcp-server and
gaia-mcp-server; the
multi-agent client in
multiagent-client-demo
drives any of these servers unchanged.
The science
Between one second and a few minutes after the Big Bang, the expanding universe ran a nuclear reactor. Its output — the primordial abundances of helium-4, deuterium, helium-3 and lithium-7 — is still measurable today, and it depends on exactly three things an agent can turn:
- the baryon density Ω_b h² (how much ordinary matter exists),
- extra relativistic species ΔN_eff (any light particle beyond the three Standard Model neutrinos — dark radiation, sterile states, axions — speeds up the expansion and changes the yields),
- the neutron lifetime τ_n (which sets the neutron/proton ratio at weak freeze-out — and whose laboratory measurements currently disagree by ~10 s, the "bottle vs beam" puzzle).

The classic stories are all here: the abundances and the CMB agree on the baryon density of the universe (concordance), helium counts the light degrees of freedom at t ~ 1 s, and the lithium-7 panel shows a famous unsolved problem — the prediction is ~3× the observed Spite plateau.
All computations run on
PRyMordial (Burns, Tait &
Valli, EPJC 84 (2024) 86), the
community Python code for precision BBN within and beyond the Standard
Model. PRyMordial is GPL-3.0 and not on PyPI, so it is not vendored
here — a setup script fetches a pinned copy into extern/ (gitignored).
Tools
| tool | what it does |
|---|---|
describe_bbn_inputs() |
parameters, measured abundances, references, the classic questions — call it first |
compute_abundances(omega_b_h2, delta_neff, tau_n_s) |
one full 63-reaction run: N_eff, Yp, D/H, ³He/H, ⁷Li/H, each with its pull (σ) against observation |
scan_baryon_density(output_dir, ...) |
abundances vs Ω_b h² → CSV (the Schramm-plot calculation) |
plot_abundance_curves(scan_file, output_dir) |
Yp, D/H, ⁷Li/H vs Ω_b h² with observed bands + the CMB baryon density |
scan_neff(output_dir, ...) |
abundances vs ΔN_eff (each point re-solves the thermal history) → CSV |
plot_neff_impact(scan_file, output_dir) |
how dark radiation shifts Yp and D/H — BBN as a particle counter |
scan_neutron_lifetime(output_dir, ...) |
Yp and D/H vs τ_n → CSV |
plot_neutron_lifetime_impact(scan_file, output_dir) |
Yp(τ_n) against the observed helium band and the discrepant bottle/beam measurements |
fit_baryon_density(scan_file, observables) |
χ² fit: the BBN-preferred Ω_b h² ± 1σ, compared with Planck — the concordance test in one number |
Predictions carry the nuclear-rate systematic (PRIMAT compilation, PRyMordial's default) in quadrature with the measurement errors — the d(p,γ)³He rate alone moves D/H by a few percent, which is a genuine, live discussion in the field, not a bug.
Two conventions worth copying into any science MCP server:
- Every tool returns
{status, files, message, metadata}(ArtifactResult). - Arrays move between tools as file paths, never through the agent's context window.
Install
conda create -n bbn-tutorial python=3.12 -y
conda activate bbn-tutorial
pip install -e ".[dev]"
python scripts/setup_prymordial.py # fetches PRyMordial + precomputes tables (~1 min)
pytest
The setup script clones a pinned PRyMordial commit into extern/PRyMordial
(override with the PRYM_DIR env var) and runs one Standard Model
computation with the save flags on, so the thermal background and weak rates
are cached — that makes parameter scans ~6× faster and lets the server run
without write access to the checkout.
Already have the spectra-tutorial env from the other servers? You can
reuse it: pip install numba numdifftools, run the setup script, then
always launch this server (and pytest) from this repo's root directory —
all tutorial servers export packages named tools and mcp_server, so
don't pip install -e several of them into one env; running from the repo
root makes the local packages win.
Run the server
python -m mcp_server --transport streamable-http --port 8003
Clients connect to http://127.0.0.1:8003/mcp. Port 8003 by convention,
so this server can run alongside spectra (8000), gaia (8001) and lattice
(8002). Stop with Ctrl+C.
To use this server from Claude Code, the Claude desktop app, Codex, Cursor,
or any other MCP client, see docs/mcp-clients.md —
a checked-in .mcp.json already wires it into Claude Code.
Questions to ask an agent
"Does the baryon density measured by Planck in the CMB correctly predict the deuterium abundance we observe in quasar spectra? Scan the baryon density, fit it from the abundances, and give me the verdict in sigma."
"If there were one extra neutrino species in the early universe, what would have happened to primordial helium, and is that allowed by the data?"
The first one becomes a genuine two-server question when this server is hosted next to a CMB-capable one (compute the damping tail there, the deuterium here).
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。