Proof
Up to 80% fewer tokens — measured, not promised.
80.6% median token reduction across 7 real repos (5,287 files), reproducible from a single command. Focused context in ~15ms. Re-run the numbers yourself.
Measured proofindex → search → context
Beyond context
AIndex does more than fetch context.
Same graph-context core, plus a full audit & analytics layer vexp doesn't ship. Based on vexp's advertised features (vexp.dev).
| Capability | AIndex | vexp |
|---|---|---|
| Graph-aware context capsule | ✓ | ✓ |
| Impact / blast-radius analysis | ✓ | ✓ |
| Persistent cross-session memory | ✓ | ✓ |
| Execution-path tracing | ✓ | ✓ |
| Dead-code detection | ✓ | — |
| Duplicate-code detection | ✓ | — |
| Code smells & audits | ✓ | — |
| Finds your most critical code | ✓ | — |
| Token-savings analytics | ✓ | — |
| Offline / air-gapped license | ✓ | — |
vexp numbers: vexp.dev and the vexp-swe-bench repo (they also report 73% SWE-bench pass@1, and support more languages/agents). Methodologies differ — a same-harness head-to-head is on our roadmap. · vexp.dev · vexp-swe-bench
| Project | Files | Tokens saved | Query |
|---|---|---|---|
| AiIndexer | 331 | 82.5% | 37 ms |
| matchr | 253 | 87% | 41 ms |
| crates | 70 | 83% | 14 ms |
| poisson | 1,380 | 81% | 73 ms |
| monsite | 192 | 78% | 2 ms |
| MyRoadTrip | 1,110 | 65% | 81 ms |
| salecast | 1,951 | 54% | 10 ms |
| Total / median | 5,287 | −80.6% | — |
Token reduction = AIndex's focused context capsule vs reading the files that contain the search hits in full, measured over each repo's most central symbols at a 32k budget. The proof also checks compression quality: top hit inclusion, top-candidate symbol coverage and top-candidate file coverage.
Compare
Pasting code into chat isn't a context strategy.
AIndex isn't another chat box. It's the local context layer that feeds your existing agent exactly the connected code each task needs.
| How you give context today | Upside | Catch | With AIndex |
|---|---|---|---|
| Copy-paste by hand | Fast for a single line | Misses everything the code connects to | Returns the connected code automatically. |
| Keyword / text search | Finds strings | Ignores how the code is wired | Ranks by the dependency graph, not just text. |
| Cloud / hosted code index | Centralized | Uploads your source to a server | Stays 100% local by default. |
| Dump whole files | Easy | Bloats the context window (20–50k tokens) | Sends a tiered capsule — typically 4–8k tokens. |
| Embedding RAG (vector search) | Semantic-ish recall | Fuzzy matches; no real call/usage graph | Ranks on the real dependency graph and shows impact — not just similarity. |
| IDE go-to-definition | Precise jumps | One symbol at a time; no task-level context | Assembles the whole task capsule in a single MCP call. |
| Other MCP code indexers | Ranked context, local | Often cap daily calls and paywall graph tools (callers, impact); their “semantic” is shallow keyword/TF-IDF | Graph navigation — references, callers, impact — included on Free. Unlimited calls. Exact, graph-resolved results. |
| Manual reference hunting | Works in your IDE | Your AI agent can’t do it — it falls back to grep (misses indirect calls, hits comments/strings) | Gives the agent Find-All-References, callers/callees and go-to-definition with exact file:line. |
Honest by default
Every number ties back to a command.
Each benchmark is generated by the binary on a local run — repo size, symbols, edges and re-index speed, with nothing massaged by hand.
ai-indexer-mcp proof --repo ~/Documents/matchr --jsonLatest run
The exact JSON, straight from the tool.
Loading benchmark data…Honest limits
What these numbers do and don't say.
These are local indexing and re-index timings on a single machine, not a hosted benchmark. Your figures depend on repo size, language mix and disk speed — run the command on your own code to see real numbers.