Zephr

32 MCP tools · 11 platform surfaces · shared registry

Intelligence · Lens

Code intelligencefor every agent.

Local lexical search and structural navigation are real today. Semantic and graph planes, plus the shared hosted runtime that will compose them, are still in design. The interaction below is illustrative, not a live hosted query.

Your question

where do we validate?

Lexical
Structural
Semantic
Graph

The four planes

Each plane is a different kind of truth.

Lexical finds the literal phrase. Structural resolves the symbol. Semantic ranks paraphrases. Graph (planned) will show what depends on this.

Four index planes over one refThe same indexed commit is read four ways — lexical, structural, semantic, graph. A query crosses the stack and stops at the plane that matched. Every answer names the plane that served it.Lexicalzephr_grepStructuralzephr_lspSemanticzephr_searchGraphzephr_searchwhere is the token checked?
  • Lexical

    Shipped
    zephr_grep

    Literal and regex indexing. The plane that never interprets, so it never hallucinates about what it found.

  • Structural

    Evolving
    zephr_lsp

    Tree-sitter and LSP symbol resolution — definitions, references, and enclosing scope.

  • Semantic

    Evolving
    zephr_search

    Code embeddings for the questions no grep answers. Embedding storage is in-database pgvector locally; the hosted semantic serving pipeline is in design, not yet serving.

  • Graph

    Planned
    zephr_search

    Import, ownership, and ref edges — designed but not yet extracted. Graph availability is planned, not live.

Every answer names the plane that served it.

LexicalShipped

Literal and regex indexing. The plane that never interprets, so it never hallucinates about what it found.

worker:apps/worker/lens-lexical-build

StructuralIn development

Tree-sitter and LSP symbol resolution — definitions, references, and enclosing scope.

worker:apps/worker/lens-structural-build

SemanticIn development

Code embeddings for the questions no grep answers. Embedding storage is in-database pgvector locally; the hosted semantic serving pipeline is in design, not yet serving.

worker:apps/worker/lens-semantic-build

GraphIn development

Import, ownership, and ref edges — designed but not yet extracted. Graph availability is planned, not live.

worker:apps/worker/lens-graph-build

Explore each engine

Pick a plane. See what it matches.

Each engine reads the same code a different way. Select one to see its tool, what it matches honestly, and where it is in the build.

Index engine explorerFour retrieval planes stacked over one codebase. Select a plane to see its tool, status, and what it matches. Every answer names the plane that served it, so the agent can verify the match rather than trusting it.Lexicalzephr_grepStructuralzephr_lspSemanticzephr_searchGraphzephr_search
  • Lexical

    Shipped
    zephr_grep

    Literal and regex indexing. The plane that never interprets, so it never hallucinates about what it found.

    example: zephr_grep "where do we validate session tokens"

  • Structural

    Evolving
    zephr_lsp

    Tree-sitter and LSP symbol resolution — definitions, references, and enclosing scope.

    example: zephr_lsp definition middleware

  • Semantic

    Evolving
    zephr_search

    Code embeddings for the questions no grep answers. Embedding storage is in-database pgvector locally; the hosted semantic serving pipeline is in design, not yet serving.

    example: zephr_search "what depends on this edge case" --hybrid

  • Graph

    Planned
    zephr_search

    Import, ownership, and ref edges — designed but not yet extracted. Graph availability is planned, not live.

    example: zephr_search "what depends on this"

Every answer names the plane that served it, so the agent can verify the match rather than trusting it.

The breakdown

One query. Four planes. One ranked, cited answer.

A breakdown of an illustrative answer. The query, the planes that contributed, and what each one gave the result.

Query

> zephr_search "where do we validate session tokens?" --hybrid

The answer

session tokens are validated in the edge middleware, never in route handlers.

apps/web/src/middleware.ts - L31-L58 - 8d13ba0

Illustrative example — not live data

How each plane contributed

Lexical

found the literal phrase across apps/web/src/middleware.ts

Structural

resolved the symbol and gave definitions, references, enclosing scope

Semantic

ranked paraphrases of the same intent from sibling files

Graph

designed to show what depends on this, on which branch it was true — graph extraction is not built yet

The other ways

You can answer the question in a worse way.

Each approach is reasonable on its own. The intended Lens answer combines distinct evidence types rather than pretending one search mode is enough. That complete hosted composition is the design target, not a live claim today.

The four ways tools usually answer

the wrong way

Embed your whole repo in a hosted vector DB

The query "what depends on this" returns semantic neighbours — not the import graph. You get plausible chunks, not the answer.

the wrong way

Use plain grep / IDE find

The query "what depends on this" returns every literal string — including false matches in tests, comments, and unrelated files.

the wrong way

Adopt a hosted dev search

A complete hosted product. Lens is a retrieval plane underneath your agents, not a complete hosted dev search. The hosted runtime is in design.

the lens way

Compose independent evidence in one answer

The design target fuses lexical, structural, semantic, and graph evidence into one cited answer. Local lexical and structural tools are real; the complete hosted composition is not live yet.

The gate

9 of 9. Measured, not claimed.

The retrieval gate is tested against nine deterministic scenarios. The IDs and verdicts are copied verbatim from docs/trust/TRUSTBENCH-EVIDENCE.md.

The nine scenarios

01Hybrid golden ordering
02Foreign-scope exclusion
03Rejected / outdated exclusion
04Semantic paraphrase
05Lexical fallback
06Contradiction abstention
07Temporal as-of gate
08Empty result clean
09Capacity-overflow abstention

How the gate runs

Deterministic, replayable, no luck.

Each scenario defines an admitted scope, a contradicting or paraphrased ref. The verdict is mechanical: a scenario that does not abstain on a contradiction is a fail.

Read the evidence

The tool surface

Three tools, one retrieval plane.

Two are single-plane - literal search, structural navigation. The third is hybrid retrieval that composes them rather than presenting separate engines as a complete hosted product.

zephr_grepLexical

Literal and regex search across admitted refs. A match is a literal one — it cannot be wrong about what it found, only about whether you wanted it.

zephr_grep "where do we validate session tokens" --refs=org/billing
zephr_lspStructural

Tree-sitter and language-server symbol resolution: definitions, references, call sites, and enclosing scope rather than string matches.

zephr_lsp definition middleware --refs=org/web
zephr_searchHybrid

Unified retrieval that combines lexical and semantic evidence into one ranked, cited answer — not a separate engine presented as a complete hosted product.

zephr_search "what depends on this edge case" --hybrid

The runtime status

Honest about what ships, what doesn't.

Every Lens runtime-status dimension with its honest state. The page does not hedge shipped as evolving or evolving as planned.

Gate layer
Tested · 9/9 scenarios
Search / grep / LSP
Real
Hybrid retrieval
Real
Hosted Lens runtime
In design
Hosted orchestration
In design
Storage
pgvector primary · Qdrant retired

A measured retrieval plane, not a complete hosted product.

Lens ships the measured gate and real local lexical and structural tools. Semantic, graph, and the shared hosted orchestration remain in design — the page says so rather than implying a complete hosted search.