mcp-loom

   ███╗   ███╗ ██████╗██████╗       ██╗      ██████╗  ██████╗ ███╗   ███╗
   ████╗ ████║██╔════╝██╔══██╗      ██║     ██╔═══██╗██╔═══██╗████╗ ████║
   ██╔████╔██║██║     ██████╔╝█████╗██║     ██║   ██║██║   ██║██╔████╔██║
   ██║╚██╔╝██║██║     ██╔═══╝ ╚════╝██║     ██║   ██║██║   ██║██║╚██╔╝██║
   ██║ ╚═╝ ██║╚██████╗██║           ███████╗╚██████╔╝╚██████╔╝██║ ╚═╝ ██║
   ╚═╝     ╚═╝ ╚═════╝╚═╝           ╚══════╝ ╚═════╝  ╚═════╝ ╚═╝     ╚═╝


   The context loom for MCP.

   One proxy in front of N servers. Oversized tool results are woven
   into queryable datasets — 48k tokens in, 2k envelope out.


   warp: results stream in         weft: the loom shuttles across
   ══════════════════════          ═════════════════════════════

       │    │    │    │
       │    │    │    │     48,377 tokens ┄┄┄┄┄┄┄┄┄> ✂ intercept
       ●━━━━┿━━━━┿━━━━┿━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
       │    ●━━━━┿━━━━┿━━━━━━━ ingest → DuckDB ━━━━━━┫
       │    │    ●━━━━┿━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┫
       │    │    │    ●━━━━ schema · stats · joins ━━┛
       │    │    │    │                              │
       ▼    ▼    ▼    ▼                              ▼
     [ raw floods, dropped ]           [ 2,104-token envelope ]
                                         ref · sample · SQL

What it does

An MCP client asks a tool a question, and the answer comes back as a ten-thousand-row JSON dump — most of which the model didn't need but now has to hold in context. Loom sits in the middle.

When a result is small, it passes through untouched. When a result is large and tabular, Loom ingests it into an embedded DuckDB and hands the model a compact envelope instead: a dataset ref, the schema, column stats, a few sample rows, join hints, and provenance. The model then queries precisely what it needs — including joins across data from different servers, which no single server can answer alone.

The economics: a 48k-token tool result becomes a ~2k-token envelope. The model spends its context on answers, not raw dumps.

Related MCP server: mcp-compressor

When loom helps — and when it won't

Loom earns its place in three situations that cover the ordinary case of real MCP usage:

• Clients with no shell. Claude Desktop, mobile, and agent platforms — the largest client surfaces — have no CLI to fall back to. MCP tools are the model's only path to the data, so interception engages by construction.

• Servers with no model-reachable CLI. Internal tools, SaaS connectors, and credentialed databases whose secrets live only in the server's env block — the bulk of real connectors. There is no gh-equivalent for the model to route around.

• Auditable analysis over persistent datasets. Loom turns "the model eyeballed a dump" into "the model ran verifiable queries over data you can still inspect" — exact SQL over provenance-stamped datasets that stay reopenable and re-queryable after the fact. This one applies everywhere, shell or no shell, and is the honest answer to "why route through loom at all."

Cross-server joins run through all three: no single CLI answers a question spanning two services' data.

Two boundaries, stated plainly:

• A shell-bearing agent will skip the MCP layer for a service it can reach directly. Give a model bash and a service with a known CLI and ambient credentials — gh against GitHub is the sharpest case — and it goes straight to the CLI, faster and cheaper, skipping loom and MCP both. That is the agent preferring a CLI to MCP, not a loom defect; loom is an MCP-layer product and will not beat gh at being gh. It is also the minority case: most servers wrap things with no model-reachable CLI, and the largest client surfaces have no shell.

• Envelope fidelity is bounded by what the downstream returns. Loom flattens faithfully what a server sends — but it cannot add fields the server never sent. The GitHub MCP server, for instance, returns a trimmed 17-field repo object (no stargazersCount, language, or forks); loom captured all 17 faithfully, and a star-ranking question was still unanswerable from that data. The bound is the downstream's, not loom's.

How it works

Loom is itself a stdio MCP server. It sits between one upstream MCP client (Claude Desktop, Claude Code, Cursor, etc.) and N downstream MCP servers that you configure:

• Spawns and supervises each configured downstream server (handshake health check, restart with backoff, delists tools and emits list_changed if a server exhausts its restart budget).

• Namespaces every downstream tool as <server>_<tool> and re-exposes the aggregated tool list upward, keeping it live via list_changed.

• Routes call_tool to the correct downstream child. Small results return unmodified — pass-through is byte-identical. Large tabular results are intercepted into DuckDB and replaced with a loom_dataset_ref envelope; any interception failure degrades to the untouched original.

• Exposes the synthetic query surface over cached datasets: loom_query (guarded read-only SQL, cross-server joins included), loom_list_datasets, loom_describe (full schema, stats, and fresh join hints for a dataset), loom_materialize (pin a derived query result as a new dataset), and loom_export (sealed csv/json export). Interception depth is capped at 1 per upstream call; loom_materialize does not consume that budget.

• Shuts down cleanly on signals or upward stdin EOF, with no orphaned child processes and the spill dir removed.

Loom intercepts both structured results and formatted text: when a result carries a structuredContent channel it ingests from that; otherwise it parses JSON-in-text, and failing that, extracts repeating records from formatted text under a strict never-lie bar (it never emits a table it isn't sure of — any uncertainty degrades to an untouched pass-through). See SPEC.md for the full normative specification.

Usage

npx mcp-loom --config loom.config.json

Point any MCP client at that command. Configuration can also be supplied via the LOOM_CONFIG environment variable instead of --config:

LOOM_CONFIG=/path/to/loom.config.json npx mcp-loom

(--config takes precedence if both are set.) See loom.config.example.json for a minimal working example.

Client recipes

Loom is launched by your MCP client the same way any stdio server is — point the client at npx -y mcp-loom --config <absolute path to loom.config.json>.

Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "loom": {
      "command": "npx",
      "args": ["-y", "mcp-loom", "--config", "/absolute/path/to/loom.config.json"]
    }
  }
}

Cursor (.cursor/mcp.json): the same mcpServers block as above.

Claude Code — add it from the CLI:

claude mcp add loom -- npx -y mcp-loom --config /absolute/path/to/loom.config.json

Use an absolute path to loom.config.json — MCP clients spawn servers with an unspecified working directory. Loom then spawns and namespaces the downstream servers listed in that config.

Config reference

Config is a single JSON file, validated with zod at startup; invalid config fails fast with a readable error listing every violated field.

If every configured server fails to start, Loom itself fails to start (fatal). If only some fail, Loom starts with a warning and the surviving servers' tools available.

Interception & querying

Large tabular tool results are intercepted and ingested into the embedded DuckDB store as queryable datasets. A tool result that exceeds tokenThreshold and carries structured data — read from the protocol's structuredContent channel when present, otherwise parsed from the text render — becomes a loom_dataset_ref envelope: a reference ID, the inferred schema, column statistics, sample rows, and join hints — keeping the token cost low. Query these datasets with the synthetic loom_query tool: it accepts read-only SQL in DuckDB dialect (single SELECT or WITH statement), classified read-only by the engine parser. If you lose a dataset reference, loom_list_datasets recovers it or lists all available datasets in the current session.

Dataset references are session-scoped: they exist only for the duration of your conversation and are automatically evicted under memory pressure (implicit query results are evicted first, preserving explicitly-ingested data longer). Sample rows in the envelope are arbitrary — DuckDB does not preserve insertion order by default — so always use an explicit ORDER BY clause when row order matters. The token estimator is a coarse magnitude dial and undercounts text containing CJK characters, so monitor actual token usage if you regularly cross tokenThreshold.

Inspecting datasets with loom_describe

Call loom_describe on a dataset ref to inspect its structure and contents in detail: full schema with per-column statistics, a 10-row sample, fresh join hints to other datasets (with value-overlap scores that rank real joins above coincidental name matches), provenance, and creation timestamp. Use this when you need to understand a dataset's shape before querying it, or to rediscover a dataset's schema after many intermediate operations.

Materializing query results with loom_materialize

Query results are cached implicitly, but if you want to keep a derived result across the session boundary or ensure it is never evicted, use loom_materialize: pass a read-only SELECT statement and get back a new pinned loom_dataset_ref envelope. The materialized dataset is never evicted under memory pressure. Important: materialized datasets are stored in their JSON text representations, so date and decimal columns become text strings. Include a CAST in your SELECT if you need date arithmetic or numeric operations on the materialized result — for example, SELECT CAST(created_at AS DATE) FROM ... to materialize a queryable date column.

Exporting datasets with loom_export (sealed boundary)

Export a cached dataset to a file with loom_export: it writes the dataset as csv or json (only these formats are supported) under the configured exportDir and returns the absolute path to the written file. This is the move for taking a result outside the session.

The sealed export boundary: The Loom engine never touches the filesystem. Export is a read-lane SELECT statement serialized to text in the Node process and written by the MCP server, not by DuckDB. This is a deliberate security choice: the engine parser enforces statement-type classification (read-only vs. write), and the engine itself is sealed against filesystem operations (COPY TO is disabled and remains disabled). The security boundary is the engine's immutable statement classification, not runtime filtering; this ensures no query can be misclassified or elevated to filesystem access at runtime.

Namespacing

Every downstream tool is re-exposed upstream as <server>_<tool>, where <server> is the config name for that server. Routing is done through an internal bidirectional map (exposed name ⇄ server + original tool name) — never by splitting the exposed name on _ — so server names containing underscores work correctly. Names that would exceed common client limits (64 chars) are deterministically truncated with a hash suffix; post-sanitization collisions get a numeric suffix. loom is a reserved server name so that a loom_-prefixed synthetic tool can never collide with a real downstream tool.

Child environment (default-deny — read this before configuring secrets)

Each downstream child gets a curated non-secret base (PATH, HOME, TMPDIR, locale — plus the Windows equivalents) plus that server's own env block from config — not a copy of the shell that launched Loom. So a secret exported in your launching shell (GITHUB_TOKEN, AWS_SECRET_ACCESS_KEY, …) is not forwarded to downstream servers, and a compromised npx -y package cannot read secrets meant for a different server.

Put per-server secrets in that server's own env block:

{ "name": "repo", "command": "npx", "args": ["-y", "@modelcontextprotocol/server-github"], "env": { "GITHUB_TOKEN": "ghp_..." } }

If one server needs an extra non-secret var from your shell, opt it in explicitly with envPassthrough (default []):

{ "name": "svc", "command": "...", "envPassthrough": ["HTTPS_PROXY", "NODE_EXTRA_CA_CERTS"] }

Boundary — least-privilege, not zero-leak. This closes the environment vector, not every vector. A var you pass through can itself carry a secret (an HTTPS_PROXY=http://user:pass@host, a CA path), and a malicious child can still reach the filesystem and network on its own. OS-level per-child isolation (containers, namespaces) is the ceiling and is out of scope for a zero-config stdio proxy — and there is deliberately no "inherit everything" flag.

Gotchas

• Child environment is default-deny, not inherited. Downstream children get a curated non-secret base plus their per-server env — the launching shell's secrets are not forwarded. Opt extra non-secret vars in with envPassthrough. Least-privilege, not zero-leak (full detail in Child environment above).

• The token threshold is a coarse dial. Interception fires on a chars/4 estimate — an order-of-magnitude trigger, not a precise token count, and it undercounts CJK text. Treat tokenThreshold as a magnitude knob, not a boundary. The estimate covers the whole result — text blocks plus serialized structuredContent when present.

• Export is sealed, and it re-serializes in Node. loom_export writes csv/json by serializing the read-lane query result to text in the Node process — the engine never writes files (COPY TO stays disabled). The cost is a JS-side serialization pass over the exported rows: bounded and well under budget for session-sized datasets (a 100k-row export measured ~0.42s), but real — an export is a copy, not a zero-cost handle.

• Dataset refs are session-scoped and evictable. Refs live only for the conversation and are evicted under memory pressure — implicit query results first, explicit ingests last; materialized refs are pinned and never evicted. Sample rows are arbitrary unless you ORDER BY.

• Downstream outputSchema is not re-exposed. Loom may replace any large result with a loom_dataset_ref envelope, so it cannot honestly advertise a downstream tool's declared output shape — a re-advertised schema would make spec-strict clients reject every intercepted result. The cost: clients don't schema-validate structured output for tools proxied through Loom; structuredContent itself still passes through untouched on small results.

Development

npm ci
npm run build    # TypeScript compile (tsc)

stdout is the protocol — all logging goes to stderr. The full specification, including the interception rules and the extraction contract, is in SPEC.md.