Mache

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 examples/ 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. Code intelligence is the flagship application of that engine — the most-developed projection, where the graph gains functions, types, cross-references, call chains, structural smell rules, and optional LSP-grade enrichment.

Mache (/mɑʃe/ mah-shay): from papier-mâché — raw material, crushed and remolded into shape.

Quick start

git clone https://github.com/agentic-research/mache.git
cd mache && task build && task install
mache init --global   # installs the keepalive HTTP daemon (:7532) + registers detected editors

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. 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 mache mount) 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. 18 MCP tools wrap the projected graph — 17 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)
Orient in an unfamiliar repo	get_overview, get_architecture, get_diagram	base	Stable
Cluster related code	get_communities	base	Beta — Louvain
Navigate by structure, not string match	find_definition, list_directory, read_file, resolve_ref, search	base	Stable
Trace the call graph, both directions	find_callers, find_callees, get_impact	base	Stable — find_callees on serve fixed in v0.18.0
Judge structural quality	find_smells — 14 rules	base¹	Stable — fan_out_skew is qualifier-aware
Edit without breaking the file	write_file — validate → format → splice, draft mode on reject	base	Stable
Types + diagnostics from a real LSP	get_type_info, get_diagnostics	+ LSP pass	Stable, optional tier
Search by meaning, not tokens	semantic_search	+ embeddings	Stable, optional tier
Check enrichment freshness	get_sheaf_status	any²	Stable
Serve across repos	(--mount NAME=PATH)	base	Stable — find_callers federates; callees per-mount
Mount as a filesystem + write back	(NFS)	base	Stable
Move a built graph between machines/CI	(mache cache push/pull/verify)	—	Stable — content-addressed, local dir or OCI
Infer a schema from data	(--infer)	—	Beta — FCA + greedy entropy

Tiers. There is no "works without ley-line-open" tier — leyline parse is 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.

For which tools read which tables, see ARCHITECTURE.md § MCP Server and § Interplay with ley-line-open.

How it works

Schema → walkers → graph → MCP / filesystem (with diagram)

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:

flowchart LR
    Source["source dir"] -->|"leyline parse<br/>(pure Go)"| Graph
    LSP["leyline lsp<br/>(LSP enrichment from<br/>ley-line-open)"] -->|"sibling .bindings.capnp<br/>(typed event log)"| BindingLog
    BindingLog -->|"ReadBindingLog"| Graph["Graph<br/>(MemoryStore or<br/>SQLiteGraph)"]
    Graph -->|"v_refs / v_defs<br/>(canonical views,<br/>fidelity poset)"| MCP["MCP tools"]
    Graph -->|"NFS server"| FS["mounted fs"]
    MCP -. "primary" .- Agent["Agent<br/>(Claude Code, etc)"]
    FS -.- Agent

Loading

1. 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.
2. Infer — schema inference (FCA + greedy entropy) discovers the natural groupings (functions/, types/, classes/)
3. 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.
4. 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.

Why this exists

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.

The graph isomorphism argument

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.

See Architecture for the full picture.

Portable cache (mache-aeb262)

Content-addressed .db bundles — push/pull/verify, OCI transport

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).

See docs/cache/phase-4-chunk-shape.md for the wire format and cloister-spec/build-cache/v1/ for the OCI transport spec.

Deployment modes

Bundle / OCI image (production) vs local dev

Mache has two supported deployment shapes:

Bundle / image (canonical production path). Mache ships its own apko + melange configs to produce a distroless OCI image (mache:0.18.0, ~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.

task image                          # → mache.tar (mache:0.18.0)
docker load -i mache.tar
docker run --rm -i mache:0.18.0 serve --stdio /path/to/source

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. (The leyline_fs FFI client is gated behind the leyline build tag and is not compiled into the image — see ADR-0006.)

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. 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 mache mount 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).

Docs

• Getting started — install + first run
• Architecture — graph backends, write pipeline, virtual directories, ley-line-open interplay
• Roadmap — what's landed, near-term, long-term
• ADRs — Architectural Decision Records
• Competitive landscape & prior art — what mache builds on, how it compares
• Example schemas — NVD, KEV, Notion, Trivy, Terraform, Markdown, LLM conversations, Go/Python/Rust/SQL source, MCP registry
• Contributing

License

Apache 2.0