Auto cross-service links
Matches what each service consumes against what others provide.
graffiti A code-graph engine for AI coding assistants
★ Zero API keys · $0 · fully offline
Stop grepping.
Stop grepping.
Start reading the graph.
Cheaper · faster · more effective
- Up to 50% fewer tokensHand the model a scoped subgraph instead of whole files — smaller, cheaper calls.
- Fewer round-tripsOne graph query instead of a dozen greps — answers land faster.
- More accurate answersThe assistant reads real structure — calls, defs, imports — not guesses.
Your AI assistant explores your codebase blind — one grep at a time. graffiti turns the whole repo into a directed knowledge graph it can query. One static binary. No API key, no cloud, no cost.
Live preview · a repo as a force-directed graph
one command:
$ graffiti . ✓ Done. 0 API calls, $0. 214 files → 1,883 nodes, 4,102 edges, 12 communities. The 3 most interesting questions your map can answer: 1) Which module is the load-bearing wall? 2) What does the auth flow touch? 3) Where are the cross-package call hotspots?
Determinism
Same repo in — byte-identical map out.
Every run sorts everything and stamps a single timestamp. Diff two builds and only the clock moves. Commit the map; review it like code.
★ Service architecture · the killer feature
Many repos.
Many repos.
One system graph.
A microservice system is N independent repos that form one product. graffiti maps each one, then discovers the edges between them — HTTP, gRPC, queues — from each service's contract.
one command — in CI or local:
$ graffiti system build ✓ System: 12 services → 38 cross-service links (2 ambiguous, 1 dangling, 4 orphan). 0 API calls, $0. $ graffiti system impact carts::GET /carts/{} ← checkout, web-bff, mobile-bff depend on it
Impact analysis
Change an endpoint → see every service that breaks.
Contract audit
Dangling consumers, orphan providers, ambiguous matches.
Each service publishes its own map
The orchestrator federates the published artifacts.
From zero to a system graph
A folder of services,
A folder of services,
mapped in three commands.
Got a folder of independent services in subdirectories? Map them into one system graph — no merge, no cloud, no config.
01
Publish each service
build then publish every subdirectory into a store at the parent folder.
02
Build the system graph
system build federates the maps and auto-discovers cross-service links.
03
Explore & guard
render the map, ask impact/query, and gate CI with audit.
in the parent folder of your services:
# parent folder; services live in ./orders, ./web, ./payments … # 1 — build + publish every service into a shared store for d in */; do graffiti build "$d" && graffiti publish "$d" --to .; done # 2 — federate + auto-discover the calls between them graffiti system build ✓ System: 3 services → 7 cross-service links (0 ambiguous, 0 dangling). 0 API calls, $0. # 3 — explore graffiti system impact orders # who breaks if orders changes graffiti system render # → .graffiti-system/system.html
Improve accuracy by dropping an openapi.json, a .proto, or a graffiti.contract.json into a service.
Vibe-chill install
Why touch a terminal
Why touch a terminal
when you have Claude Code?
Paste this one prompt into your Claude Code session. Say "y" when it asks for permission. Done.
No Go toolchain, no build flags, no config hunt — Claude downloads the binary, maps your repo, and wires up the integration, asking before each command.
you paste in Claude Code:
Install graffiti by amazopic for me. Download the right static binary for my OS/arch from the latest release at github.com/amazopic/graffiti (or build it from source with `make build` if Go is available), put it on my PATH as `graffiti`, and verify with `graffiti version`. Then run `graffiti .` at my repo root to build the map, run `graffiti init --hook` to wire graffiti into Claude Code, and finally open `.graffiti/map.html` so I can see the graph. Ask before each step.
↳ Just say y (yes) at every permission prompt — Claude will run each command one by one.
Editor's Note
An AI coding assistant is only as good as what it can see.
graffiti is the map that should have been there.
By the numbers
7
languages parsed0
API calls, ever$0
dollars to run1
static binary (~10 MB)23
documented languages100%
byte-deterministic, offline
Token cost
One question.
One question.
A fraction of the tokens.
Ask “where is the cart fetched and served?” — here's what your assistant has to read to answer it: grepping blind versus one scoped graph query.
grep + read files≈ 4,200 tokens
greps, opens ~7 candidate files, reads them whole
graffiti query≈ 1,900 tokens
one scoped subgraph — defs, callers, callees
≈ 55% fewer tokens
Illustrative — actual savings vary by repo and task. graffiti query caps each answer at a soft ~2,000-token budget, so context stays lean and cheap.
One command,
three artifacts.
graffiti . writes everything into .graffiti/.
map.json
The graph itself — nodes, edges, communities, schema-checked.
MAP.md
A human-readable digest of the most important structure.
map.html
A single self-contained, offline, interactive force-directed graph.
map.html · interactive viewer
- 2D / 3D toggle — hover lifts a node and its neighbours
- search nodes · click to copy file:line
- sector zones · client / tests / external toggles
- resizable project → directory → file tree
A side-by-side
vs. grep,
vs. grep,
vs. the cloud
| Capability | Grep / cloud RAG | graffiti |
|---|---|---|
| Directed graph: calls, defs, imports | no | yes |
| Runs fully offline, no account | no | yes |
| Ongoing cost | $ per token | $0 |
| LLM-free, token-budgeted retrieval | no | yes |
| MCP server built in | no | yes |
| Byte-deterministic output | no | yes |
| Interactive offline graph viewer | no | yes |
| What it needs | cloud + API keys | one binary |
Get going
Install in 30 seconds
One static binary does it all: build the map, query it, serve it over MCP, and wire itself into Claude Code.
01
Get the binary
One curl — or make build from source.
02
Map your repo
Run graffiti . at the repo root.
03
Wire up Claude Code
graffiti init --hook — and your assistant reads the graph.
★ Or in Claude Code — easiest
Have Claude Code open? Paste this one prompt and you're done — no terminal needed.
01
you Install graffiti by amazopic for me…
macOS / Linux / WSL
# 1 — get the binary (verifies SHA256) curl -fsSL https://raw.githubusercontent.com/amazopic/graffiti/main/scripts/install.sh | sh # 2 — map your repo graffiti . # 3 — look at the graph open .graffiti/map.html
The installer picks the right static binary for your OS/arch and verifies its SHA256. Prefer source? make build produces the same ~10 MB CGO-free binary.
Claude Code integration
graffiti init --hook # skill + CLAUDE.md + grep→query nudge graffiti serve # MCP server over stdio graffiti query "auth flow" # LLM-free scoped subgraph
Multi-repo
Frontend and backend,
Frontend and backend,
one graph.
Lay separate repos side by side and query across them — without merging.
Assert cross-project edges, validate they resolve, then query the whole federation.
$ graffiti link ../frontend ../backend ✓ Linked 2 projects. 7 cross-project links. 0 API calls, $0. $ graffiti query --workspace "where is the cart fetched and served"
Q & A
Frequently