io.github.brainfoods-studio/llmgraph

LLMGRAPH

Every call graph has blind spots. Ours tells you where they are.

LLMGRAPH inks only the calls it can prove. The ones it can't see, it counts and names in plain words, so your agent sees the blind spots before it edits.

I built LLMGRAPH as the truth layer for coding agents, and the truth includes the blind spots: agents guess; LLMGRAPH proves what it can and names what it can't. In a blind test of 50 questions on three open-source repos, no confident answer was wrong.

Two functions. One name. Your agent can't tell. We can.

grep reads names. LLMGRAPH follows calls. What it can prove, it proves; what it can't, it tells you, down to the line where it stopped.

zod ships two copies of addIssueToContext. A text search mixes them together:

$ rg -n addIssueToContext      # zod v3.23.8
146 lines, from both copies, mixed together

LLMGRAPH separates them, and shows its work:

llmgraph · who_calls addIssueToContext (src/helpers/parseUtil.ts:72): 71 in a set
  ⚠︎ type checker starting: answers use our own reading for now; 71 calls unchecked
  ◇ within_these 71 calls -> one of 1 (closed: repo only) [c] w1
  ⚠︎ plus 7 dynamic calls nearby that could reach anything w2
  – ruled out: 71 same-name calls (proven: other target) [c] w3
  assumes [c]: our own reading, not confirmed by the checker
(lines trimmed)

71 look-alikes ruled out, and it shows why. And where the type checker hasn't confirmed yet, it says so.

Who's holding the other end?

Before your agent pulls a thread, it should know. In requests, change super_len(o) to super_len(o, strict), and ask check_my_changes:

llmgraph · check_my_changes: 1 file changed, 16 calls break
  16 breaks: misses "strict" w1
  ⚠︎ breaks if it runs super_len src/requests/models.py:526 PreparedRequest.prepare_body (within_these, one of 1, closed: repo only) [c] w2
  ⚠︎ breaks if it runs super_len src/requests/models.py:575 PreparedRequest.prepare_content_length (within_these, one of 1, closed: repo only) [c] w3
  ⚠︎ breaks if it runs super_len tests/test_utils.py:62 TestSuperLen.test_io_streams (within_these, one of 1, closed: repo only) [c] w4
  … 13 more test call sites
  ⚠︎ plus 9 dynamic calls nearby that could reach super_len w18
(lines trimmed)

16 broken calls named, exactly the real ones. A blind tester's text search found the same 16, no more and no fewer. And the 9 dynamic calls it can't follow are marked, not ignored.

Then the type checker catches up. Ask again a few seconds later:

llmgraph · who_calls super_len (src/requests/utils.py:135): 16 certain
  ✓ certain src/requests/models.py:526 PreparedRequest.prepare_body w33
  +1 more (all certain, same file) w34
  ✓ certain tests/test_utils.py:62 TestSuperLen.test_io_streams w35
  +13 more (all certain, same file) w36
  ⚠ plus 9 dynamic calls nearby that could reach anything w37
(lines trimmed)

From "our own reading" to 16 certain, on the released program. Its labels never say more than it has proven, and they upgrade the moment it can.

The five labels

Other tools sketch your codebase from rumour and fill the blind spots with confidence. LLMGRAPH works in two colours: ink for what it has proven, a plain warning for what it hasn't, and never one dressed up as the other.

Answers carry no numbers about chance; they say what is ruled out.

Measured on our test projects

Callers are confirmed by the type checker in under a second after your agent asks (Python), and check_my_changes answers in under a second once LLMGRAPH has caught up with a save. On names that also appear in unrelated code, it named no wrong callers where a plain text search returned hundreds of wrong lines, and the agent read less than half as many tokens; on clean, unique names there is no saving.

What your agent can ask

Your code is read and analysed on your machine, and we never receive your code, paths, symbol or repo names, queries or answers. Checked on the released program: it contacts only beta.llmgraph.in. Usage counts and crash reports are on by default and can be switched off (npx -y llmgraph-mcp telemetry off); the licence check is always on. Privacy page.

Install

Claude Code (plugin):

claude plugin marketplace add brainfoods-studio/llmgraph-agent-plugin
claude plugin install llmgraph@llmgraph

The plugin offers LLMGRAPH's answer before your agent searches for a function or class name, and never blocks the search. When your agent is about to finish, it stops it once if the saved edits left calls that are certain to break, and lists them; it never blocks on anything less, and npx -y llmgraph-mcp hook off turns both off.

Any MCP agent: add the server npx -y llmgraph-mcp (on Windows some agents need cmd /c npx -y llmgraph-mcp), sign in once with Google, and your agent asks. Setup guides for Claude Code, Antigravity, Cursor, Codex, Gemini CLI, Windsurf and OpenCode; tested by us on Claude Code and Antigravity.

Today's limits: Windows x64 · Python and TypeScript/JavaScript (other languages are reported as cannot_see, not counted as nothing) · repos up to the plan's line limit · invite-only beta, request an invite · one plan for individuals with a trial, prices.

This repo holds the open setup files under MIT; the LLMGRAPH engine is closed source and runs on your machine under the EULA. Brainfoods Studio, India · founder@llmgraph.in · security@llmgraph.in