What it does
Blast Scope scores what a shell command would do before it runs. Instead of matching command syntax against a blocklist, it works out the command's real target, observes that target with a read-only probe, and returns a structured risk score with supporting evidence. The same command can receive different scores depending on what it would affect. It covers filesystem, git, Docker, pip and uv, and SQL operations. The scoring is advisory and never blocks a command. For critical commands it can capture an undo snapshot first.
It suits anyone running AI coding agents who wants a check on destructive shell commands before execution. The publisher describes a PreToolUse hook that can run this check inside Claude Code. The assess_command tool scores a possibly chained command and returns a score, severity, rationale, evidence, recoverability, affected nodes and a per-segment breakdown. The index_project tool builds the project's dependency graph on demand. The publisher reports a 58-case in-repo corpus and results from the SABER benchmark, with a 0.58% false-positive rate on about 1,725 safe commands.
The fastest install is zero-install through uvx, and Claude Code can be wired up with one line that adds both the MCP tools and the advisory hook. For development, the server can also run from a source checkout. Live probes need a working environment for the target, so when a probe cannot run, the tool returns a labelled estimate rather than a silent guess.
Limitations and requirements
Live probes need a Docker daemon, a database driver or credentials. Without them the tool returns a labelled estimate. Scoring is advisory only and does not block commands. Network exfiltration and persistence are out of scope by design. The publisher's benchmark recall differs across categories.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
io.github.Atharva-Jayappa/blast-scope
Contextual blast-radius scoring for shell commands an AI agent is about to run
v0.6.1As shown on registry.modelcontextprotocol.io when we last read it