Mache projects structured data into a navigable graph. Point a declarative schema at any structured source — JSON documents, SQLite databases, source code — and mache exposes the projection as MCP tools or a mounted filesystem, so an agent (or you) can explore the topology instead of grepping flat files.
One engine, many projections. The schemas in project CVE feeds (NVD, KEV), Notion exports, Trivy scan results, Terraform, Markdown, LLM conversation logs, and audit trails with the same machinery that projects Go or Rust source. — the most-developed projection, where the graph gains functions, types, cross-references, call chains, structural smell rules, and optional LSP-grade enrichment.
IDE-style navigation for structured data — code, JSON, YAML. Jump to definitions, find callers, follow references. Available as an MCP server or a mounted folder.
2026-10-27
Sao
46
Công cụ
—
Thời gian hoạt động
—
examples/
Code intelligence is the flagship application of that engine
Mache (/mɑʃe/ mah-shay): from papier-mâché — raw material, crushed and remolded into shape.
That's the 30-second path for the flagship application — code intelligence on a Go codebase. mache init --global installs a per-user supervisor (launchd on macOS, systemd --user on Linux) that keeps the shared mache HTTP daemon alive on localhost:7532 and registers it with Claude Code and any detected editors — no terminal to babysit. Then mache init (no flag) inside a project records what that project serves.
Two things new users hit:
HTTP is the canonical transport. One shared daemon serves every project, routing per session via the MCP roots protocol. --stdio exists only as an escape hatch for CI / sandboxes / headless agents and is never registered for editor use (see ADR-0022).
ley-line-open is a hard requirement, not an add-on. Since v0.18.0 mache has no in-process parser: leyline parse produces the _ast database that every source projection reads. task install provisions the pinned binary from a checkout; mache install does the same from a release asset or package-manager install, where there is no Taskfile. Point mache at a directory or a pre-built .db — both route through leyline; a pre-built .db just skips the parse step and can carry LSP/embedding enrichment.
To project non-code data instead, hand mache serve (or the root mount form, mache <mountpoint>) a schema with --schema and point it at your JSON or SQLite source — the example schemas cover NVD, KEV, Notion, Trivy, Terraform, and more.
For the full first-run flow — source choice (directory vs .db vs live hot-swap), the --stdio escape hatch, --scope, Claude Desktop, mount as filesystem, write-back, schema inference, troubleshooting — see GETTING-STARTED.md.
Code intelligence: what it gives an agent
The code projection is where the engine is deepest. 19 MCP tools wrap the projected graph — 18 read-surface plus write_file.
What you get
Tools
Tier
Status
Parse source into a graph
—
required
Stable — 27 of 28 languages (all but cue: no grammar)
Tiers. There is no "works without ley-line-open" tier — leyline parseis the parser.
base — any source projection. leyline parse → _ast → pure-Go ASTWalker, from a directory or a pre-built .db.
+ LSP pass — leyline's LSP tier, auto-spawned on demand (pass a file=) or pre-baked into a .db. Drives the real language server (rust-analyzer, gopls, …); verified end-to-end on Rust and Go. Same primitive Serena is built on, but as an enrichment tier over the _ast base rather than the foundation.
+ embeddings — a ley-line-open-built .db carrying embedding vectors.
¹ 5 of the 14 rules need an enriched .db. ² Reports daemon cache state; returns {available: false} when unreachable, so it's safe to call anywhere.
Internal machinery — canonical v_refs/v_defs views (ADR-0013), the capnp event-log readthrough, snapshot memoization, the e2e tool/flamegraph harness — lives in ARCHITECTURE.md and CHANGELOG.md.
Every projection follows the same shape: a schema declares the topology, walkers extract nodes and edges from the source (JSONPath for JSON, direct SQL for SQLite, AST queries for code), and the resulting graph is served over MCP or mounted as a filesystem. The source-code path — the most developed — looks like this:
Parse — leyline parse (from ley-line-open) turns source into an _ast database. mache reads it pure-Go via the ASTWalker; in-process CGO tree-sitter was removed in ADR-0012 step 4.
Link — cross-reference extraction builds a call graph from identifiers and imports. When the LSP pass from ley-line-open has run, refs flow through a sibling .bindings.capnp typed event log (per ADR-0013) rather than SQL columns — the wire format is the cross-runtime contract.
Project — the graph is exposed as MCP tools (primary) or a mounted filesystem (optional)
The graph is the same on either path; MCP and the filesystem are two ways to talk to it.
Agents operate without topology. They see flat files, grep for strings, build a mental model, forget it next turn, rebuild it. The structure is in the data — functions call other functions, types reference types, configs depend on configs — but nothing exposes it.
Mache does. Point it at data, it figures out the shape. Source code gets parsed by leyline parse. JSON and YAML get walked. Schema inference discovers the natural groupings without config. The agent can then explore the topology directly: follow call chains, find definitions, read context, write back.
Built for agents first. The design choices — stable node paths across edits, identity-preserving write-back — exist because agents need to reference things reliably across turns. The outputs are human-discernible because the representations are filesystems and SQL, but the topology is the point.
Both structured data and filesystems are graphs. Your JSON object has nodes and edges (containment). Your filesystem has nodes and edges (parent-child). They're isomorphic.
Operating systems never formalized this mapping. Mache does:
SQL is the graph operator — queries define projections from one topology to another
Schema defines topology — the formal specification of how source nodes map to filesystem nodes
The filesystem exposes traversal primitives: cd traverses an edge, ls enumerates children, cat reads node data
ADR-0011 takes this further: every navigable thing in mache (path, token, SHA, range, record, ref) is a Pointer; the graph is a network of pointers; mache resolves them on demand.
Library callers can produce either a raw leyline database or a Mache
schema-projected database without invoking the mache CLI:
import (
"github.com/agentic-research/mache/build"
"github.com/agentic-research/mache/schema"
)
// Resolve a bundled preset or a schema file relative to the project.
resolved, err := schema.Resolve("go", projectDir)
if err != nil {
return err
}
// Use the resolved topology directly...
if err := build.ParseWithSchema(projectDir, outputDB, resolved.Topology); err != nil {
return err
}
// ...or let build preserve preset metadata used by the grammar guard.
return build.ParseWithSchemaRef(projectDir, outputDB, "go", projectDir)
build.Parse writes leyline's raw parse shape. ParseWithSchema and
ParseWithSchemaRef additionally run Mache's Engine + ASTWalker
projection, including the registered env:, mod:, and gomod: reference
extractors. All paths invoke the pinned leyline executable and keep Mache
CGO-free. Address-reference extraction uses leyline's root-relative source IDs,
so files below the source root participate in the same graph as root files.
Running the pinned leyline (mache-19326d)
mache pins an exact ley-line-open release and caches it version-namespaced under ~/.mache/bin, which is not on PATH. So a bare leyline in your shell runs whatever copy PATH happens to hold — often an older one — against .db files mache parsed with the pin. Both are named leyline; neither warns. Since leyline v0.13.0 shipped leyline cdc enable / leyline cdc gc, invoking leyline directly is a real workflow, so this is a real skew.
Two subcommands make the version question unaskable instead of answered-by-PATH-convention:
# Absolute path of the pinned binary, and nothing else on stdout.
mache leyline path
# Compose it — version-correct by construction.
$(mache leyline path) cdc enable --db ./mache.db
# Or proxy straight through: args, stdin/stdout/stderr and exit status forwarded.
mache leyline exec -- cdc enable --db ./mache.db
mache leyline exec cdc gc --db ./mache.db
Both resolve through mache's pin-checked chokepoint, so a non-pinned leyline earlier on PATH is reported on stderr and ignored rather than silently used. Neither mutates PATH, your shell rc, or the unversioned ~/.mache/bin/leyline path (which stays never-written — see internal/leyline/binary_cache_path.go).
MACHE_NO_LEYLINE=1 still disables provisioning; MACHE_LEYLINE_BINARY still overrides the pin for LLO development.
Portable cache (mache-aeb262)
mache cache push/pulls the projected .db as a content-addressed bundle so CI / new dev machines / agents don't re-parse a million lines on every cold start.
# Emit a portable bundle from a built db.
mache cache push --db ./mache.db ./cache-out
# Restore a fresh db from a bundle.
mache cache pull --out-db ./restored.db ./cache-out
# Push to a remote OCI registry (build-cache/v1 transport).
mache cache push --db ./mache.db ./cache-out \
--remote https://cache.example.com --scope myrepo/abc123 --tag latest
# Pull from a remote registry into a local cache dir, then restore.
mache cache pull --out-db ./restored.db ./cache-in \
--remote https://cache.example.com --scope myrepo/abc123 --ref latest
# CI-friendly: assert a bundle is intact + verifiable, without restoring.
mache cache verify \
--remote https://cache.example.com --scope myrepo/abc123 --ref latest
When the source db has an _ast table, chunks carry the AST node rows too — pull restores both _source AND _ast, so the restored db is queryable without re-parsing. When _ast is absent, chunks are raw source bytes (Phase 1 fallback).
Bundle / image (canonical production path). Mache ships its own
apko +
melange configs to produce a
distroless OCI image (mache:0.22.0, ~33MB, x86_64 + aarch64). This is the
unit that a cluster orchestrator (e.g. cloister) deploys; inside the
bundle, mache speaks to a co-located ley-line daemon over a UDS socket
and is unreachable except via the orchestrator-mediated wire.
Given a fixed melange.rsa signing key and pinned toolchain, the build is
reproducible — same input git tree, same image hash. task image
auto-generates a dev keypair when one is missing (APK signatures will
differ across freshly-generated keys), so for byte-stable artifacts in CI
inject a fixed keypair from a secret. The melange recipe builds with
CGO_ENABLED=0 — mache is pure Go since ADR-0012
step 4 removed in-process tree-sitter. mache links no external C libraries; it
consumes leyline as a subprocess/daemon over its UDS socket, not via FFI.
On release, mache also self-publishes a separate leyline-bundled
multi-arch image to ghcr.io/agentic-research/mache (debian-slim +
libsqlite3, since leyline links sqlite; the distroless apko image above
stays the local-dev build). This image gets the ley-line-open-paired path
with no runtime fetch.
Its ENTRYPOINT is ["mache", "serve"], so docker run args append to
mache serve — do not repeat serve, or you get mache serve serve:
docker run --rm -i ghcr.io/agentic-research/mache:v0.22.0 --stdio /source
Mache declares its own source via server.json's
packages[].oci entry, tag-pinned — see
ARCHITECTURE.md § OCI distribution
for the full framing.
Local / dev path — running mache serve or the root mount form
(mache <mountpoint>) directly on
your machine. Useful for laptop work, debugging, and writing schemas. In
this mode mache may auto-discover or auto-download a leyline binary
(legacy code path; the bundle ships everything, so this only kicks in for
non-bundle invocations). Set MACHE_NO_LEYLINE=1 to disable auto-download
in CI or when leyline-open hasn't published a release for your platform.
Exposing this mode externally requires a reverse proxy in front for auth;
mache itself does not implement perimeter auth (the bundle gets it from
the orchestrator).
IDE-style navigation for structured data — code, JSON, YAML. Jump to definitions, find callers, follow references. Available as an MCP server or a mounted folder.
Chưa có danh sách công cụ. Máy chủ từ xa sẽ có danh sách sau lần kiểm tra tình trạng đầu tiên.
Câu hỏi thường gặp
Câu hỏi
Làm thế nào để cài đặt Mache?
Sao chép cấu hình stdio cho Cursor, Claude Desktop, Claude Code hoặc VS Code trên trang này. Cấu hình dùng bản cài đặt của nhà phát hành, không phải proxy của PlugMCP.
Mache có cần khóa API không?
Bản ghi này không liệt kê biến môi trường bắt buộc nào.
Mache dùng giao thức truyền tải nào?
Mache chạy cục bộ qua stdio.
Mache được kiểm tra lần cuối khi nào?
Kiểm tra lần cuối: 2026-10-27. IDE-style navigation for structured data — code, JSON, YAML. Jump to definitions, find callers, follow references. Available as an MCP server or a mounted folder.
