blastradius

Source repository · Issues · PyPI package

Temporal code knowledge graph with blast-radius impact scoring, semantic symbol search, and git-history-aware dependency analysis — for AI-assisted development.

Point it at any project — Python, JavaScript/TypeScript, Go, Ruby, Rust, Java, PHP, and more — and get:

• A persistent SQLite graph store at <repo>/.blastradius/index.db — incremental, queryable, temporal
• A blastradius.json dependency index written directly into your repo (preserved for backward compatibility)
• Per-file blast-radius scores (how many files break if this one changes), including historical as-of queries
• A symbolindex.json symbol map so AI can find any function/class without scanning every file
• Hybrid semantic search over symbols: natural-language queries fused with keyword + graph expansion
• Git history backfill — temporal graph from commit history without touching the working tree
• Ten ways to consume the data: CLI, markdown report, MCP server (10 tools), pre-commit hook, CLAUDE.md injection
• An interactive visualization UI (2D/3D graphs, dependency matrix, treemap)

No project build step or npm required. SQLite is stdlib; the default installation includes the official Python MCP SDK and its runtime dependencies.

Install

Install uv first, then install the published tool into its own isolated environment:

uv tool install blastradius-cli
blastradius --help

The package is blastradius-cli; the executable is blastradius. This repository maintains and publishes the PyPI package. You do not need to activate a virtual environment or add blastradius to the project you want to analyze. Python 3.10+ is required. MCP is included by default; no MCP extra is needed. See uv's tool guide for details.

If your shell cannot find blastradius, run uv tool update-shell and open a new terminal. uv tool dir --bin shows the executable directory; use command -v blastradius on macOS/Linux or Get-Command blastradius in PowerShell to check which executable your shell finds. Remove an older installation with its original installer if it shadows the uv-managed command.

Run without a persistent installation

uvx --from blastradius-cli blastradius --help
uvx --from blastradius-cli blastradius analyze ./myapp

uvx is shorthand for uv tool run. Keep --from blastradius-cli: the package name differs from the command name. For regular use, MCP clients, and git hooks, prefer the persistent installation above.

Upgrade or uninstall

uv tool upgrade blastradius-cli
uv tool list
uv tool uninstall blastradius-cli

Upgrades retain the source and version constraints used for installation.

Install from this repository

PyPI contains the latest published release, not unreleased commits. To install the default branch directly from this repository (requires Git):

uv tool install git+https://github.com/IWasZ3r0Cool/blast-radius.git

Or install a checkout, after selecting the branch or commit you want:

git clone https://github.com/IWasZ3r0Cool/blast-radius.git
cd blast-radius
uv tool install .

This installs a snapshot of the checkout. Re-run uv tool install --reinstall . after changing it, or use uv tool install --editable . to intentionally track local edits. Keep the checkout available for editable installs. Source installs replace the same blastradius-cli tool; they are not a second executable.

Develop and test blastradius itself

Inside this repository, use the project environment rather than the installed tool:

uv sync --extra dev
uv run blastradius --help
uv run pytest
uv run ruff check blastradius tests
uv build

The packaging test builds an sdist and wheel, installs the wheel with uv tool install into temporary tool/bin directories, and exercises the installed command outside the checkout. It does not modify your personal tool installation.

Quickstart

After installation, run commands from the repository you want to analyze. Replace the example file and release tag below with ones that exist in that repository.

cd ./myapp

# Build the dependency index (also writes to .blastradius/index.db)
blastradius analyze .

# Build the symbol index (where every function and class lives)
blastradius symbols .

# See blast radius for a file before touching it
blastradius impact src/auth.py

# Search symbols with natural language (no embedding endpoint needed — FTS fallback)
blastradius search "validate auth token"

# See what changed since a release tag
blastradius changed-since v1.2.0

# Launch the visualization UI
blastradius serve --viz --repo .

Leave the server running and visit http://localhost:8080/ in your browser. Use the HTTP URL, not the HTML file directly; the UI loads your index from /graph.

Commands

blastradius analyze

blastradius analyze [REPO_PATH] [--output PATH] [--watch]

Analyzes the repo and writes blastradius.json to the repo root. Detects 12+ languages automatically.

blastradius symbols

blastradius symbols [REPO_PATH] [--output PATH] [--inline] [--index PATH]
                  [--claude-md] [--claude-md-path PATH] [--all-symbols]

Builds a symbol index — a map of every function, class, struct, and type to its exact file and line number. Lets AI tools (and humans) find any symbol in one lookup instead of scanning the entire repo.

Modes:

Both --inline and --claude-md can be combined in a single run.

Options:

Examples:

# Standalone symbol index
blastradius symbols ./myapp

# Embed into blastradius.json (one file for blast radius + symbols)
blastradius symbols ./myapp --inline

# Write CLAUDE.md summary so Claude Code loads symbols automatically
blastradius symbols ./myapp --claude-md

# All three at once
blastradius symbols ./myapp --inline --claude-md

# Re-generate when code changes
blastradius symbols ./myapp --inline --claude-md

Why it matters: Claude Code and other AI tools normally scan every file to find a function definition. With a symbol index, Claude can load one file, do an O(1) lookup, and open only the relevant file — cutting token usage 60–90% on symbol-location tasks.

CLAUDE.md injection is opt-in because it increases base context size on every prompt. Use it when symbol lookups are frequent in your workflow; skip it for simple tasks where the overhead outweighs the benefit.

blastradius impact

blastradius impact FILE [--index PATH] [--out FILE] [--json] [--as-of REF]

Shows the blast-radius impact for a specific file: direct dependents, transitive dependents, blast score, and risk level.

Impact: src/auth.py
Blast Score: 8.5  (2 direct · 7 transitive)  [HIGH]

Direct dependents (2)
  src/api.py
  src/middleware.py

Transitive dependents (5 additional)
  src/main.py  ← src/api.py
  src/app.py   ← src/middleware.py
  ...

Risk: HIGH — affects 7/42 files (16.7% of codebase)

Blast score formula: direct + (0.5 × transitive)

blastradius search

blastradius search QUERY [--k N] [--as-of REF] [--db PATH] [--json]

Hybrid semantic + keyword + graph symbol search. Finds relevant functions and classes without knowing their exact names.

Retrieval signals fused with Reciprocal Rank Fusion (RRF):

1. Semantic KNN — embedding similarity (requires blastradius-cli[semantic] + a configured endpoint)
2. FTS5 keyword — full-text search over symbol names, signatures, and docstrings
3. Graph expansion — structurally adjacent symbols from dependent/dependency files

Degrades gracefully: if no embedding endpoint is configured, falls back to FTS + graph (no crash, no config needed).

# Keyword + graph search (no embedding setup required)
blastradius search "validate auth token"

# Full semantic search (requires BLASTRADIUS_EMBEDDING_* env vars)
BLASTRADIUS_EMBEDDING_ENDPOINT=http://localhost:11434 \
BLASTRADIUS_EMBEDDING_MODEL=nomic-embed-text \
BLASTRADIUS_EMBEDDING_DIMS=768 \
blastradius search "validate auth token"

# Historical search — symbols visible at a release tag
blastradius search "token validation" --as-of v1.2.0

blastradius history

blastradius history [REPO_PATH] [--since REF] [--max-commits N] [--json]

Backfills temporal graph data from git history — without any working-tree checkouts. Reads blobs via git cat-file --batch and stamps first_seen_commit / last_seen_commit on every file, edge, and symbol.

Run this once after initial setup to enable --as-of queries across the full history.

# Backfill up to 1000 commits (default)
blastradius history .

# Backfill only recent history
blastradius history . --since v1.0.0 --max-commits 200

blastradius changed-since

blastradius changed-since REF [--repo PATH] [--db PATH] [--json]

Lists files and dependency edges added or removed since a commit, branch, or tag.

$ blastradius changed-since v1.2.0
Changes since v1.2.0:

  Added files (2):
    + src/payments/stripe.py
    + src/payments/webhook.py

  Removed edges (1):
    - src/api.py → src/auth_v1.py  [imports]

blastradius db

blastradius db status [--db PATH] [--json]
blastradius db migrate [--db PATH]

Manages the SQLite store at <repo>/.blastradius/index.db.

$ blastradius db status
schema_version      : 2
repo_root           : /Users/alice/myapp
last_indexed_commit : a3f2e1c8
active_files        : 142
active_edges        : 387
active_symbols      : 1204
embedding_model     : nomic-embed-text
embedding_dims      : 768
vec_symbols         : enabled

db migrate applies any pending schema migrations automatically (also runs on every blastradius analyze).

blastradius serve

blastradius serve --viz [--repo PATH] [--port PORT] [--watch]
blastradius serve --mcp [--repo PATH]

--viz starts a local HTTP server for the interactive visualization UI (5 modes: 2D force graph, 3D network, dependency matrix, treemap, infrastructure graph). Visit the printed URL in your browser; keep the terminal process running. The installed tool includes the HTML, so no source checkout is needed. Use --port if port 8080 is already occupied.

--mcp starts a stdio MCP server that exposes blastradius tools directly to Claude and other MCP clients.

For a project-specific server, use either launcher:

blastradius serve --mcp --repo /path/to/project
uvx --from blastradius-cli blastradius serve --mcp --repo /path/to/project

The SDK supports modern MCP 2026-07-28 requests and legacy initialization with 2024-11-05, 2025-03-26, 2025-06-18, and 2025-11-25. Modern clients use server/discover and per-request metadata; discovery lists modern revisions. Legacy clients continue to use initialize followed by notifications/initialized. See the MCP compatibility specification.

--repo sets a fixed default repository; it does not analyze it automatically. Explicit index_path, symbol_index_path, and db_path arguments take precedence. Relative artifact and analysis paths are based on --repo, or on the startup working directory when omitted. Without --repo, indexes are discovered by walking up from that directory. With --repo, a missing index never falls back to a parent or unrelated repository. Analyzing another repository does not change this default. File queries prefer exact repo-relative or absolute paths; shortened paths must identify a unique file.

For a graph exported outside its repository, keep the export selected with index_path and configure --repo /path/to/project when querying absolute source paths. get_impact and get_dependencies check exact node matches relative to the graph's directory and the configured repository. Without --repo, repository context comes from the nearest graph or database above the startup directory, or that directory itself if neither exists. Conflicting matches are reported as ambiguous. If no context matches, use a repo-relative file path or configure --repo; absolute paths are never guessed by suffix. The selected graph's data is never replaced, and its meta.root basename is not an absolute-location hint.

Tool arguments are validated without string-to-number coercion. semantic_search.k must be a positive integer; graph_query.depth must be a nonnegative integer (0 returns the starting node only). Malformed calls, invalid arguments, and unknown tools return protocol errors. Missing files/indexes and invalid Git refs return failed tool results with actionable messages; an unknown symbol is a successful lookup with found: false.

MCP uses stdin/stdout, not a browser URL. Diagnostics go to stderr. If a client cannot connect, check the executable path, Python requirement, and client stderr. If a tool cannot find its index, run blastradius analyze /path/to/project (or blastradius symbols /path/to/project for a standalone symbol index) and check --repo or its explicit artifact argument. Closing stdin shuts down the server. Cancellation prevents queued work from starting and suppresses late responses; an already-running synchronous write finishes safely, without rollback.

MCP tools:

Claude Code project MCP config (.mcp.json):

{
  "mcpServers": {
    "blastradius": {
      "command": "/absolute/path/to/blastradius",
      "args": ["serve", "--mcp", "--repo", "/absolute/path/to/project"]
    }
  }
}

Replace the command with the full path under uv tool dir --bin (append blastradius, or blastradius.exe on Windows). See the setup walkthrough below.

blastradius lookup

blastradius lookup SYMBOL [--index PATH] [--json]

Finds where a function, class, struct, or other symbol is defined. Queries the SQLite DB (same source as search), falling back to symbolindex.json if no DB is present. Prints 5 lines of source context around the definition.

$ blastradius lookup compute_blast_radius
blastradius/impact.py:6  compute_blast_radius  (function)

  >    6 | def compute_blast_radius(nodes: list[dict], links: list[dict]) -> dict[str, dict]:
         7 |     """..."""
         8 |     ...

$ blastradius lookup AuthService
src/auth.py:44  AuthService  (class)  methods: login, logout, refresh

  >   44 | class AuthService:
        45 |     def login(self, ...):

If a symbol isn't found, it's likely a third-party import — only symbols defined in the repo are indexed.

blastradius dependencies

blastradius dependencies FILE [--index PATH] [--json]

Shows what a file imports and what imports it, plus its blast score.

$ blastradius dependencies src/auth.py
File: src/auth.py  (blast score: 8.5)

Imports (3):
  src/db.py
  src/config.py
  src/utils.py

Imported by (2):
  src/api.py
  src/middleware.py

blastradius high-blast

blastradius high-blast [--threshold N] [--index PATH] [--json]

Lists all files whose blast score exceeds the threshold, sorted by score descending. Useful for identifying the riskiest files before a refactor.

$ blastradius high-blast --threshold 5
Files with blast score ≥ 5.0  (3 found)

  13.0  src/db.py          (12d / 2t)
   8.5  src/auth.py        (3d / 7t)
   5.5  src/config.py      (5d / 1t)

d = direct dependents · t = transitive dependents

blastradius symbol-blast

blastradius symbol-blast FILE [--json]

Per-export blast radius for a file. Lists every exported symbol with the count and exact file paths of importers that reference it by name. Useful when a file exports many symbols and you only need to touch one — lets you confirm the others are safe.

$ blastradius symbol-blast lib/db/schema.ts

Symbol-level blast radius: lib/db/schema.ts

  5 exported symbol(s), 12 importer(s)

  userSchema  (const, line 8)  →  8 user(s)
    app/api/users/route.ts
    app/api/profile/route.ts
    ...
  legacySchema  (const, line 42)  →  1 user(s)
    scripts/migrate.ts
  sessionSchema  (const, line 67)  →  3 user(s)
    ...

userSchema touches 8 routes; legacySchema touches only 1 — safe to change in isolation.

blastradius install-hook

blastradius install-hook [--repo PATH] [--threshold N] [--strict] [--remove]

Installs a git pre-commit hook that warns when staged files exceed the blast score threshold.

Using blastradius in another repo with Claude

Three workflows, ordered by automation level.

Workflow 1 — MCP server (recommended for active coding)

Claude gets symbol lookup, dependency, and impact tools it calls automatically. No extra prompting needed.

One-time setup:

cd /your/other/repo
blastradius analyze .
blastradius symbols .

Register the MCP server with Claude Code using claude mcp add. Use --scope project for this repo, or --scope user for all your repos. These are the scopes in Claude Code's MCP documentation.

# Project-scoped (stored in .mcp.json)
claude mcp add --scope project blastradius -- /path/to/blastradius serve --mcp --repo /path/to/project

# User-scoped (available in all your repos)
claude mcp add --scope user blastradius -- /path/to/blastradius serve --mcp

Run uv tool dir --bin, append /blastradius (\blastradius.exe on Windows), and substitute that absolute path above. On macOS/Linux, this can be done directly:

claude mcp add --scope user blastradius -- "$(uv tool dir --bin)/blastradius" serve --mcp

Verify it registered:

claude mcp list

Note: A GUI or MCP client may have a different PATH from your terminal. An absolute executable path avoids relying on its shell configuration. Do not commit machine-specific executable or repository paths in shared .mcp.json without adapting them for your team. A user-scoped server without --repo relies on the client's launch directory; use explicit artifact paths when querying a different repository, or configure a project-specific server.

Claude now has all 10 MCP tools available in every session. When it needs to find processPayment, it calls lookup_symbol("processPayment") and gets src/billing.py:142 back in one shot — no file scanning. When it needs to find code that validates auth tokens without knowing the exact name, it calls semantic_search("validate auth token").

Keep the index fresh:

# Auto-rebuild the dependency index (requires the watch extra; leave running)
blastradius analyze . --watch

# Rebuild the standalone symbol index after edits, in another terminal
blastradius symbols .

symbols does not have a --watch flag.

Workflow 2 — CLAUDE.md injection (best for repos you revisit often)

Symbol table is embedded in CLAUDE.md so it loads into every session automatically — no tool call needed at all.

cd /your/other/repo
blastradius symbols . --claude-md

This upserts a symbolindex code fence into CLAUDE.md. Every Claude Code session in that repo loads it at startup. Claude can answer "where is X defined?" from context alone with zero tool calls.

Tradeoff: adds ~500–2000 tokens to every prompt depending on repo size. Worth it for repos where symbol lookups are frequent; skip it for repos where you mostly write new code.

Keep it fresh:

# Re-run after significant refactors
blastradius analyze . && blastradius symbols . --claude-md

Workflow 3 — Hybrid (large repos)

For large repos where the --claude-md section would be too large, use the MCP server for lookups and add a short hint to CLAUDE.md so Claude reaches for the tool first:

blastradius analyze .
blastradius symbols .

Then add to CLAUDE.md:

## Blastradius
Symbol index: `symbolindex.json` — use the `lookup_symbol` MCP tool before grepping for any function or class.
Dependency index: `blastradius.json` — use `get_impact` before modifying high-blast files.

This costs almost no tokens but primes Claude to use the index rather than defaulting to grep.

CLI quick reference for human use

The same data available via MCP is also accessible directly from the terminal:

# Find where a symbol is defined (shows code snippet)
blastradius lookup MyClassName
blastradius lookup process_payment --json

# Show what a file imports and what depends on it
blastradius dependencies src/auth.py

# List the riskiest files to change
blastradius high-blast --threshold 5

# Blast-radius report before touching a file
blastradius impact src/auth.py

# Per-export blast radius — which importers use each exported symbol
blastradius symbol-blast lib/db/schema.ts

Which workflow to pick

Supported Languages

Output schemas

blastradius.json

{
  "meta": {
    "root": "myapp/",
    "total_files": 60,
    "total_loc": 4085,
    "languages": ["python", "javascript"]
  },
  "nodes": [
    {
      "id": "src/auth.py",
      "type": "module",
      "language": "python",
      "layer": "backend",
      "loc": 142,
      "imports": ["src/db.py"],
      "imported_by": ["src/api.py", "src/middleware.py"],
      "direct_dependents": 2,
      "transitive_dependents": 7,
      "blast_score": 5.5,
      "symbols": [
        { "name": "verify_token", "line": 18, "kind": "function", "exported": true },
        { "name": "AuthService",  "line": 44, "kind": "class",    "exported": true,
          "methods": ["login", "logout", "refresh"] }
      ]
    }
  ],
  "links": [
    { "source": "src/api.py", "target": "src/auth.py", "weight": 1, "kind": "imports" }
  ]
}

The symbols field is only present when blastradius symbols --inline has been run.

symbolindex.json

{
  "meta": {
    "generated": "2026-05-21",
    "repo": "myapp/",
    "total_symbols": 312
  },
  "symbols": {
    "verify_token": [
      {
        "file": "src/auth.py",
        "line": 18,
        "kind": "function",
        "exported": true,
        "doc": "Verify a JWT and return the decoded payload."
      }
    ],
    "AuthService": [
      {
        "file": "src/auth.py",
        "line": 44,
        "kind": "class",
        "exported": true,
        "methods": ["login", "logout", "refresh"]
      }
    ]
  },
  "file_symbols": {
    "src/auth.py": [
      { "name": "verify_token", "line": 18, "kind": "function", "exported": true },
      { "name": "AuthService",  "line": 44, "kind": "class",    "exported": true,
        "methods": ["login", "logout", "refresh"] }
    ]
  }
}

Lookup patterns:

• "Where is verify_token defined?" → symbols["verify_token"][0].file + .line — O(1)
• "What symbols live in src/auth.py?" → file_symbols["src/auth.py"] — O(1)
• "What's the blast radius of changing verify_token?" → cross-reference blastradius.json via the file

CLAUDE.md symbol section

When --claude-md is used, a compact section is upserted into CLAUDE.md bounded by HTML comment markers so re-runs update in place:

<!-- blastradius-symbols-start -->
## Symbol Index
_Generated by blastradius. Update: `blastradius symbols --claude-md`_

```symbolindex
src/auth.py: verify_token:fn:18 AuthService:cls:44[login,logout,refresh]
src/db.py: connect:fn:12 query:fn:28 close:fn:55

Format per symbol: `name:kind_abbr:line[methods...]`
Kind abbreviations: `fn` function · `cls` class · `st` struct · `en` enum · `tr` trait · `if` interface · `ty` type · `co` const

---

## AI workflow comparison

| Task | Without blastradius | With symbolindex.json |
|------|-------------------|----------------------|
| Find where `process_payment` is defined | Grep / scan ~200 files | Load 1 file, O(1) lookup |
| Understand blast radius of a change | Manual tracing | `blastradius impact <file>` |
| Load only relevant context | Full repo scan | File + line from symbol map |
| Estimated token savings | baseline | **60–90% on symbol tasks** |

---

## Optional dependencies

The MCP SDK, AnyIO, and JSON Schema validation are required runtime dependencies
installed automatically. The following features remain optional:

| Package | Purpose | Install |
|---------|---------|---------|
| `sqlite-vec` | Semantic vector search in `blastradius search` and `semantic_search` MCP tool | `uv tool install --force 'blastradius-cli[semantic]'` |
| `watchdog` | `--watch` file change detection | `uv tool install --force 'blastradius-cli[watch]'` |
| `PyYAML` | Better Docker Compose / CI YAML parsing | `uv tool install --force 'blastradius-cli[yaml]'` |
| `tomli` | Rust `Cargo.toml` on Python < 3.11 | `uv tool install --force 'blastradius-cli[toml]'` |

These commands replace the installed tool's configuration with the requested
extras; they do not accumulate extras from previous commands. Combine everything
you need in one specification:

```bash
uv tool install --force 'blastradius-cli[semantic,watch,yaml,toml]'

The commands above install the published package. To retain a source checkout, run uv tool install --force '.[semantic,watch,yaml,toml]' from that checkout. Manage dependencies through uv tool, not by modifying its environment directly. For contributor tests, use uv sync --extra dev --extra semantic instead.

Semantic search configuration

Semantic search requires a self-hosted embedding endpoint (Ollama, LM Studio, llama.cpp server, or any OpenAI-compatible /v1/embeddings API) and the sqlite-vec extension.

uv tool install --force 'blastradius-cli[semantic]'

# Export in the shell that launches blastradius (or its MCP client)
export BLASTRADIUS_EMBEDDING_ENDPOINT=http://localhost:11434
export BLASTRADIUS_EMBEDDING_MODEL=nomic-embed-text
export BLASTRADIUS_EMBEDDING_DIMS=768

# Work in the repository to be indexed and searched
cd ./myapp
blastradius analyze .

# Search
blastradius search "validate JWT token"

The installed command does not automatically load .env files.

Without these env vars, blastradius search and the semantic_search MCP tool fall back to FTS5 keyword + graph search — no crash, no config required.

Requirements

• uv for the recommended tool installation
• Python 3.10+
• A modern browser (for --viz mode)

License

Apache 2.0 — free to use and build on; attribution required in derivative works and documentation. Copyright 2026 David Scheiderman.