petromcp
A local-first MCP server for reading petroleum data formats like LAS and DLIS, designed for teams whose data cannot leave their machine. It provides read-only tools for well log analysis, comparison, and unit conversion with strict path allowlisting and no network access.
README
petromcp
Read well logs with Claude without the data ever leaving your machine.
petromcp is an MCP server for petroleum data formats, built for teams whose files legally cannot be uploaded to a cloud service. No telemetry, no phone-home, no automatic updates, and a default-deny path allowlist that refuses to open anything you have not explicitly permitted. LAS today; DLIS, SEG-Y, and pump cards next.
If you can upload your data somewhere, you have more options than this. If you can't, this was written for you.
What this gives you
LLM hosts cannot read binary or semi-structured petroleum formats. petromcp
wraps the established open-source parsers — lasio and dlisio today, with
segyio queued for a later slice — and exposes them as MCP tools, so
you can have a conversation with your data instead of copy-pasting curve
values into chat.
What "local-first" means here, concretely
Not a posture. Four things you can verify in the source:
- Default deny. A fresh install can read nothing. Access is granted per directory, and every tool routes through one validator that resolves symlinks before checking, so a link inside an allowed directory cannot reach outside it. There is no environment variable that widens the allowlist and no tool that changes it at runtime.
- No network, declared. petromcp opens no outbound connections. All five
tools ship
openWorldHint: falseandreadOnlyHint: true, so your host can verify that claim rather than take it. - An audit trail. Every tool call is logged with a timestamp, the tool name, and the resolved path.
- Bounded output. Curve reads are capped and report
downsampledandoriginal_count, so a 20,000-point log cannot silently dump into a context window.
The threat model, and what does not count as a vulnerability, are in SECURITY.md. Read docs/DATA_PRIVACY.md before pointing this at real data — it is authoritative, and if the code contradicts it the code is the bug.
Install
Requires Python 3.10+ and uv.
Add petromcp to your MCP host's config — no clone, no build:
{
"mcpServers": {
"petromcp": {
"command": "uvx",
"args": ["petroleum-mcp", "serve"]
}
}
}
On macOS that file is
~/Library/Application Support/Claude/claude_desktop_config.json. Restart
the host afterwards. macOS notes and troubleshooting:
docs/INSTALL.md.
To work on petromcp rather than just use it:
git clone https://github.com/ameyxd/petromcp
cd petromcp
make setup
make install-claude
Configure
By default petromcp can read nothing. Tell it which directories are fair game:
uvx petroleum-mcp config init
uvx petroleum-mcp config add-path ~/petroleum/wells
Or, if you want to try it without your own data, generate the synthetic sample from a checkout and allowlist that:
make generate
uvx petroleum-mcp config add-path "$(pwd)/examples/sample_data"
Restart your MCP host after editing the allowlist — it is read once at startup.
Use
Open a new conversation and ask, in plain language:
What's wrong with this well log? /path/to/well.las
Compare these two wells: /path/to/A.las and /path/to/B.las
Convert 1500 psi to kPa.
Claude picks the right tool, reads the file through petromcp, and answers.
Worked examples, with real tool output rather than prose:
- QC a well log — finds a planted density gap, a washout, and a gamma ray spike
- Compare two wells — depth overlap, a missing curve, and a unit mismatch that is invisible in one file
- Unit conversion
- Read a DLIS file — logical files, frames, and what happens when a channel name is ambiguous
Every value in those documents is generated by calling the tools; CI fails if they drift.
Tools
| Tool | What it does |
|---|---|
read_las_file |
Header-level summary of a LAS file |
summarize_las_curves |
Per-curve min, max, mean, stddev, gap percentage |
read_las_curve |
Depths and values for one curve, with sampling cap |
compare_well_logs |
Common curves, depth overlap, unit consistency, flags |
convert_units |
ft<->m, psi<->kPa, psi<->bar, bbl<->m3, degF<->degC, mD<->m2 |
list_supported_units |
Every convertible pair with its physical quantity |
read_dlis_file |
DLIS structure: logical files, frames, index types |
list_dlis_channels |
Every DLIS channel with its frame, units, and length |
read_dlis_channel |
One DLIS channel's values, with a sampling cap |
qc_a_well_log prompt |
Walks Claude through a standard QC pass |
Every tool is read-only and opens no network connection, and declares that in its MCP annotations. Full reference: docs/TOOLS_REFERENCE.md.
DLIS files hold several logging runs, each with several frames, so a channel
name is unique only within a frame. read_dlis_channel refuses an ambiguous
name and lists the candidates rather than guessing, because the values differ.
SEG-Y and pump card support land in subsequent releases.
Status
v0.7 ships the LAS and DLIS slices, a comparison tool, a units utility, and config-management CLI subcommands. The remaining formats are tracked in SPEC_petromcp.md. The non-goals list there is real; read it before filing feature requests.
Release history: CHANGELOG.md. Security policy and threat model: SECURITY.md.
License
MIT.
<!--
The official MCP registry verifies that whoever publishes
io.github.ameyxd/petromcp also controls the PyPI package, by looking for this
marker in the package's README. Removing it makes registry publishing fail with
a 400; make release-check guards it.
-->
mcp-name: io.github.ameyxd/petromcp
Built by Amey Ambade. I write about AI systems in industries where the data can't leave the building, at writing.heyamey.com.
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。