codinflow

Code intelligence for Claude Code and AI agents. Ask a JS/TS repo precise questions — who calls this, what breaks if I change it, what runs when this route is hit — and get answers from the TypeScript compiler's resolved call graph, not a text search. grep floods an agent with 40 ambiguous matches; codinflow returns the one correct, type-resolved answer, so it gets there in fewer reads and fewer tokens.

Entirely local: no install, no account, no server.

Give your agent the tools

npx skills add Prajeevan/codinflow   # via skills.sh — Claude Code, Cursor, Cline, Copilot, …
npx codinflow skill install          # or install the skill directly if you have the CLI
npx codinflow .                      # analyze the current repo (caches a graph in .codinflow/)

See it as a visual map (for humans)

npx codinflow --ui          # analyze the current folder and open the visual canvas
npx codinflow --ui ./my-app

This analyzes the folder, starts a local server (default http://127.0.0.1:9338) and opens the interactive canvas in your browser — files, functions, routes, data stores and how they connect. Entirely local: no upload, no token, no hosted service. Re-running reuses the cached graph when your code hasn't changed.

Analyze to a file / pipe

npx codinflow ./my-app --out graph.json
bunx codinflow ./my-app --out graph.json
pnpm dlx codinflow ./my-app --json | jq .stats

A public GitHub repository, by owner/repo shorthand or full URL:

npx codinflow honojs/hono --out hono.graph.json

Upload to a self-hosted CodinFlow API so it shows up in a shared canvas (the local --ui needs none of this):

npx codinflow ./my-app \
  --api "$CODINFLOW_API" \
  --token "$CODINFLOW_TOKEN"

Options

| Flag | Meaning | | --- | --- | | --repository-id <id> | Name it in the UI (default: folder or repo name) | | --commit-sha <sha> | Label this snapshot (default: current git HEAD) | | --branch <name> | Branch to clone, for a GitHub URL | | --out <file> | Write the graph JSON to this file | | --api <url> | Upload to a self-hosted CodinFlow API | | --token <token> | Bearer token for --api (or set CODINFLOW_TOKEN) | | --fn <name> | Symbol for query (also the first operand of describe/impact) | | --file <path> | For impact: a whole file instead of a symbol | | --route <route> | For trace (also its first operand) | | --depth <n> | Max walk depth for impact (12) / trace (8) | | --json | Machine-readable output, on every query verb | | --refresh | Re-analyze before answering — guaranteed fresh |

The first argument is a local path (use ./) or a GitHub owner/repo / URL. A leading analyze verb is optional: codinflow analyze ./my-app also works.

Cloning a remote repo runs git clone --depth 1 and nothing else — repositories are parsed, never installed or built.

Ask the codebase questions (for humans and AI agents)

analyze on a local folder caches a warm graph in .codinflow/. Every query verb reads it — and, crucially, reports how stale it is versus the working tree, so an answer is never silently out of date.

codinflow map ./my-app             # orientation: routes, hotspot files, boundaries, env vars
codinflow query --fn getRouter ./my-app   # who calls/imports it — and behind which guards
codinflow describe createOrder ./my-app   # one symbol's full story: signature, traits,
                                          #   callers, callees, reads/writes/throws/external
codinflow impact getRouter ./my-app       # blast radius: transitive callers → routes,
                                          #   importing files (type invalidation), test files
codinflow impact --file src/db.ts ./my-app
codinflow trace "POST /api/orders" ./my-app  # middleware order + guarded call tree
                                             #   + db/external/env touches
codinflow status ./my-app          # is the cached graph current?

All of it is type-resolved, and guards are verbatim source — impact will tell you a caller only runs if (!isValidShopifyOrder(order)). Add --json to any verb for machine-readable output, --refresh to re-analyze first.

For query, --output is a comma list of: calls, usedBy, importedBy, reads, writes, throws, external (default calls,usedBy,importedBy).

AI agents: npx codinflow skill install writes a .claude/skills/codinflow skill that teaches Claude Code (and compatible agents) the whole flow — map to orient, describe/query while reading, impact before editing, trace when debugging an endpoint.

Staleness is first-class

Every query answer carries a verdict, computed by hashing the working tree against the cached graph (cheap — no compiler):

• fresh — nothing changed since the graph; answer at full confidence.
• stale-unaffected — the graph is old, but no file this answer depends on changed, so it still holds.
• stale-affected — a file this answer depends on changed; re-run with --refresh.

This is what makes the cache trustworthy for an AI agent: type-resolved "who calls / imports this" in one structured call, and it tells you when it might be wrong. --json emits the report plus the staleness block for programmatic use.

Requirements

Node.js >= 18. typescript is installed automatically as a dependency.

License

MIT