lodeline

mcp
Guvenlik Denetimi
Gecti
Health Gecti
  • License — License: Apache-2.0
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Community trust — 15 GitHub stars
Code Gecti
  • Code scan — Scanned 12 files during light audit, no dangerous patterns found
Permissions Gecti
  • Permissions — No dangerous permissions requested

Bu listing icin henuz AI raporu yok.

SUMMARY

Open-source config scanner for AI coding agents: scores the permissions, hooks, sandbox and MCP settings of Claude Code, Codex, Cursor and Gemini CLI. Runs locally and measures by default, with an optional fleet server and SIEM feed. Rust.

README.md

Lodeline

A config scanner for AI coding agents. Claude Code, Codex, and Cursor can run
whatever their config allows. Lodeline scores what that is, in one command,
before something does.

Formerly Sigil. Upgrading an existing install? See the
migration guide.

License
macOS
Linux
Windows
Rust
Status

A coding agent doesn't ask before it runs a hook, launches an MCP server, or
skips its sandbox. Those decisions live in config files you never open. Lodeline
reads them and tells you how exposed you are.

curl --proto '=https' --tlsv1.2 -fsSL https://raw.githubusercontent.com/Ju571nK/lodeline/main/install.sh | sh
lodeline scan
Lodeline scan — CRITICAL (7.5)
2 tools assessed · 5 findings

TOOL         SCOPE               SCORE  BUCKET    TOP FINDINGS
claude-code  user-global           7.5  critical  no_sandbox, broad_matcher (+1 more)
cursor       application:cursor    2.5  medium    mcp_server_local_command, no_sandbox

5 tools not configured: antigravity, claude-desktop, codex, continue-dev, gemini

How to reduce
  no_sandbox                     The agent runs with no sandbox boundary — enable the tool's sandbox or limit what it can run.
  broad_matcher                  A broad hook matcher catches most or all tool calls — scope it to specific tools.
  destructive_in_inline_command  A hook runs a destructive shell command inline — review and remove it.
  mcp_server_local_command       An MCP server auto-launches a local command — review or remove it, or run it sandboxed.

Run `lodeline scan --json` for full detail.

Your EDR sees the command that ran. Lodeline sees the permission that let it run.

Runs on your machine · no account · nothing leaves the box · macOS / Linux / Windows · Rust · Apache-2.0

Try the 60-second tour →


What it catches

The misconfigurations that turn a helpful agent into a foothold:

  • No sandbox — the agent runs with full reach into your host. no_sandbox
  • Wildcard hooks — a .* PreToolUse matcher that lets any tool call through, or a hook that runs a destructive inline command. broad_matcher · destructive_in_inline_command
  • Local-command MCP servers — an mcp.json server set to auto-launch a shell or binary. mcp_server_local_command
  • Empty deny lists — permissions that look configured but block nothing. permissions_deny_empty
  • Prompt-injection in instruction files — directives planted to steer the agent off-task.

Across the agents you actually run: Claude Code, Codex, Cursor, Gemini CLI, Antigravity, Continue.dev, Claude Desktop.

It re-scores the moment a config changes — a clean repo at 0 / low jumps to
critical the instant a risky hook lands.


Why your linter and EDR miss this

  • A linter / SAST reads your source. It says nothing about what your agent is allowed to do.
  • An EDR flags the process after it launches. By then the permission that allowed it already existed.
  • Lodeline reads that permission surface (sandbox, hooks, MCP, instruction files) and scores it before anything runs.

It measures. It doesn't block by default. Blocking exists, but it is opt-in, so your agent keeps working.


Running agents on several machines?

If you've got Claude Code and Codex grinding away unattended on a rack of Mac
minis, each box has its own posture and you can't eyeball them all. The same scan
rolls up across machines, so one view tells you which host is riskiest and why —
optionally shipped to your SIEM and a fleet dashboard.

That's the fleet side: a client agent on every machine, hash-anchored events
over mTLS to a central server, signed policy pushed back down, read-only MCP for
operators. It's there when you need it and invisible when you don't.

→ Fleet setup · lodeline-manager dashboard · architecture


Install

# personal (default): lodeline + lodeline-mcp + lodeline-hook, no server
curl --proto '=https' --tlsv1.2 -fsSL https://raw.githubusercontent.com/Ju571nK/lodeline/main/install.sh | sh

# fleet: adds lodeline-sender + lodeline-server + lodeline-sign
curl --proto '=https' --tlsv1.2 -fsSL https://raw.githubusercontent.com/Ju571nK/lodeline/main/install.sh | LODELINE_PROFILE=fleet sh

Linux .deb/.rpm and Windows .zip are on the releases page.
Every release ships a SHA256SUMS (the installer verifies it) and a build-provenance attestation.
Full guide: docs/install-personal.md.


Ask your agent about its own risk

lodeline-mcp exposes the score over plain MCP, so an AI client can read its own
posture and explain it:

  • lodeline-check — this host only (the default a coding agent registers): my_risk, my_guard_detail, my_findings
  • lodeline-fleet — read-only fleet view for operators (GET only)

No vendor plugin, no write path by construction.


How it works

A small Rust agent watches the config and posture files on each machine, hashes
them, and emits JSONL posture events. Locally that's all you need. For a fleet,
lodeline-sender ships those events over mTLS to lodeline-server, which feeds your
SIEM and the optional lodeline-manager dashboard and pushes signed policy back.

The workspace has twelve crates with clear roles (lodeline-agent, lodeline-sender,
lodeline-server, lodeline-mcp, lodeline-hook, lodeline-signer, lodeline-core,
and others). lodeline-core is forbid(unsafe_code). Built on rustls, ed25519,
blake3, notify, and axum.


Status

Lodeline is alpha, built and maintained by one person. It measures posture and
does not block
by default. SECURITY.md is honest that there's
no SLA. The Linux runtime is a working foundation with rough edges (watch limits,
coverage signaling) still on the roadmap. Use it to see your
exposure today; don't treat it as a managed enterprise control yet.

Apache-2.0 · issues and discussions welcome · landing

Yorumlar (0)

Sonuc bulunamadi