Skip to content

README.md

Latest commit

 

History

History
876 lines (635 loc) · 43.9 KB

README.md

File metadata and controls

876 lines (635 loc) · 43.9 KB

hyperdb-mcp

Note: This crate is AI-assisted but human-directed — much of the code was written by AI coding assistants under close review, with the design and engineering trade-offs decided by an experienced developer. The pre-1.0 (0.x) line may still undergo large breaking changes; the public API won't settle until the 1.0.0 release.

An MCP (Model Context Protocol) server that turns the Hyper columnar database into an instant SQL analytics engine. Data flows in from other MCP plugins or files, lands in Hyper automatically, and becomes queryable with SQL — no setup, no schema files, no database management.

Built on the pure-Rust hyperdb-api crate for maximum performance: 22M+ rows/sec inserts, 18M+ rows/sec queries, constant memory for billion-row results.

Why

LLMs are powerful at reasoning but cannot natively crunch millions of rows. This plugin bridges that gap: another MCP tool produces data, the LLM passes it to hyperdb-mcp, Hyper ingests it and makes it SQL-queryable, the LLM runs analytical SQL, and results come back as JSON. Optionally export to CSV, Parquet, Apache Iceberg, Arrow IPC, or .hyper (opens directly in Tableau Desktop).

Queryable Memory for AI

Unlike flat-text memory systems that store blobs and retrieve by similarity search, HyperDB gives LLMs structured, queryable long-term memory. The persistent database survives across sessions — anything the LLM stores there can be JOINed, filtered, aggregated, and reasoned over with full SQL in any future conversation.

This means an LLM can:

  • Accumulate knowledge over time — store reference tables, project decisions, user preferences, learned facts
  • Cross-reference across sessions — JOIN today's analysis against historical data from last week
  • Answer complex recall questions — "Which projects had budget overruns in Q1?" is a SQL query, not a fuzzy text search
  • Build on prior work — load yesterday's cleaned dataset and extend it without re-processing from scratch
  • Maintain structured context — store relationship graphs, timelines, or decision logs as proper tables with typed columns

The ephemeral database is scratch space (think: a whiteboard). The persistent database is long-term memory (think: a filing cabinet you can query). Multiple AI clients sharing the same daemon see the same persistent data — so Claude Code, Cursor, and VS Code Copilot can all read from and contribute to the same knowledge base.

Features

  • Zero setup — HyperProcess auto-starts the Hyper server
  • Shared hyperd daemon — one Hyper process per user, shared across all MCP clients (Claude Code, Cursor, VS Code, etc.) for reduced memory overhead and concurrent access to the same persistent databases
  • Queryable long-term memory — persistent database survives across sessions; LLMs can store, recall, JOIN, and aggregate structured knowledge over time — not just retrieve text blobs, but reason over them with SQL
  • Any data in — JSON, CSV, Parquet, Arrow IPC, Apache Iceberg; schema inferred or exact
  • SQL at scale — thousands to billions of rows
  • Data out — export to CSV, Parquet, Apache Iceberg, Arrow IPC, or .hyper (Tableau Desktop-ready)
  • One-shot queries — query_file("/tmp/sales.csv", "SELECT ...") — single call, zero management
  • Cross-session continuity — load multiple tables, JOIN across them, persist across sessions; pick up exactly where you left off
  • Read-only safe mode — --read-only flag for safe deployment
  • Schema resources — auto-discover table schemas via resources/list
  • Guided prompts — analyze-table, compare-tables, data-quality, suggest-queries
  • Inline charts — bar/line/scatter/histogram as PNG or SVG
  • Incremental ingest — watch_directory monitors for .ready sentinel files
  • Performance telemetry — every response includes throughput stats
  • Smart schema inference — exact (Arrow/Parquet), structural (JSON), heuristic (CSV) with full-file numeric widening
  • Pre-ingest file inspection — inspect_file dry-runs the same inference without touching Hyper so LLMs can build safe schema overrides in one shot
  • Partial schema overrides — supply just the columns you want to correct (e.g. {"population":"BIGINT"}) — the rest keep their inferred type
  • Rich resource surface — workspace readme, per-table JSON and CSV samples, and one JSON + one CSV resource per table so LLMs can orient themselves via resources/list without any tool calls
  • Saved queries — register named read-only SQL with save_query; each query becomes hyper://queries/{name}/definition (metadata) + hyper://queries/{name}/result (live re-run). Persisted in the persistent attachment, session-only when --ephemeral-only
  • Live resource-update notifications — MCP clients can resources/subscribe to any hyper://... URI; the server fires notifications/resources/updated after every ingest, DDL, watcher event, or saved-query mutation

Installation

From npm

Requirement: Node.js v21 or later. Earlier versions ship an older npx whose argument parsing is incompatible with the npx -y hyperdb-mcp invocation in the MCP config below. If you're on an older Node, see Upgrading Node.js with nvm below.

npm install -g hyperdb-mcp

The npm package bundles both the hyperdb-mcp binary and the hyperd database server — no additional setup required.

Upgrading Node.js with nvm

nvm (Node Version Manager) makes it easy to install and switch between Node.js versions.

macOS / Linux (nvm-sh/nvm):

# install nvm if you don't have it
curl -o- https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.1/install.sh | bash

# install and use the latest LTS (>= 21)
nvm install --lts
nvm use --lts
node --version    # should report v22.x.x or newer

Windows (coreybutler/nvm-windows): download the installer, then in a new shell:

nvm install lts
nvm use lts
node --version

After upgrading, restart your MCP client so it picks up the new Node binary on PATH.

Building from Source

cd hyper-api-rust
cargo build --release -p hyperdb-mcp

The binary is at target/release/hyperdb-mcp.

When building from source the hyperd executable is not bundled, so you'll need to provide one. The easiest path is the companion hyperdb-bootstrap CLI, which downloads a matching pinned hyperd for your platform:

cargo install hyperdb-bootstrap
hyperdb-bootstrap download                # installs into ./.hyperd/current/
export HYPERD_PATH="$PWD/.hyperd/current" # or pass via your MCP config

hyperdb-bootstrap also has a library API if you'd rather wire the download into your own build script — see its README. If you already have hyperd elsewhere (Tableau Hyper API for C++/Python/Java ships one), point HYPERD_PATH at it or add it to your PATH.

MCP Client Configuration

Each AI tool reads MCP server config from a different file but uses the same JSON shape. The base config block using npx (recommended):

{
  "mcpServers": {
    "HyperDB": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "hyperdb-mcp"]
    }
  }
}

Or if you built from source:

{
  "mcpServers": {
    "HyperDB": {
      "type": "stdio",
      "command": "/path/to/hyperdb-mcp",
      "env": {
        "HYPERD_PATH": "/path/to/hyperd"
      }
    }
  }
}

By default, persistent storage lives at the platform data dir (~/Library/Application Support/hyperdb/workspace.hyper on macOS, ~/.local/share/hyperdb/workspace.hyper on Linux, %APPDATA%\hyperdb\workspace.hyper on Windows). To use a custom path:

"args": ["--persistent-db", "/path/to/my-project.hyper"]

Multiple MCP clients can point at the same persistent file simultaneously — they all connect through the shared hyperd daemon and use Hyper's MVCC transaction isolation. See Operating Modes below.

Claude Code / AI Suite

Create or edit ~/.claude/.mcp.json (global) or .mcp.json in the project root (project-scoped). Use the base config block above.

After adding the config:

  1. Start a new Claude Code session. You'll be prompted to approve the server on first use.
  2. Auto-approve tools (optional): Add "mcp__HyperDB__*" to the permissions.allow array in ~/.claude/settings.json.

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows). Use the base config block above.

Cursor

Edit ~/.cursor/mcp.json (global) or .cursor/mcp.json (project root). Use the base config block above.

Other MCP Clients

Any tool that supports the MCP stdio transport can use this server. Point it at the hyperdb-mcp binary and set HYPERD_PATH in the environment.

Operating Modes

Each session has two databases: an ephemeral primary (scratch space — always created fresh per session, deleted on exit) and a persistent database (queryable long-term memory — stored at the platform-default location or a path you supply, survives indefinitely). Unqualified SQL targets the ephemeral primary; the persistent database is reachable as the "persistent" alias.

Hyper engine

Mode Flag Behavior
Shared daemon (default) (none) One hyperd process per user, shared across all MCP clients. The first client auto-spawns the daemon; subsequent clients discover and reuse it. The daemon stays resident by default (idle shutdown is opt-in — see below), so the next client connects instantly instead of waiting for a fresh hyperd to start. A client built from a newer hyperdb-mcp version transparently takes over (stops and replaces) an older running daemon.
Private hyperd --no-daemon Each MCP client spawns its own hyperd (legacy behavior, one per session).

The shared daemon is the bigger win for users running multiple AI clients (Claude Code + Cursor + VS Code) — they all share one Hyper engine instead of spawning three.

Database storage

Mode Flag Behavior
Default (none) Ephemeral primary in $TMPDIR/hyperdb-mcp-<pid>-<n>/scratch.hyper + persistent attachment at the platform data dir (e.g. ~/Library/Application Support/hyperdb/workspace.hyper on macOS).
Custom persistent path --persistent-db <PATH> Same as default but the persistent file lives at <PATH>. The deprecated --workspace <PATH> is accepted as an alias with a stderr warning.
Ephemeral-only --ephemeral-only No persistent attachment; the session has only the ephemeral primary plus any user-attached databases via attach_database. Saved queries fall back to in-memory storage and disappear when the session ends.

HYPERDB_PERSISTENT_DB overrides the default persistent path the same way --persistent-db does.

Working with both databases

Tool calls default to the ephemeral primary — that's the LLM's scratch space for exploratory work that doesn't need to outlive the session. To store data in long-term memory (the persistent database), there are two ways to reach it:

1. Per-tool database parameter (preferred for ergonomic LLM workflows):

// Save a useful table to the persistent database
load_data({ table: "customers", data: "[...]", persist: true })
//   ↑ shorthand for `database: "persistent"`

// Query from persistent
query({ sql: "SELECT * FROM customers", database: "persistent" })

// Inspect persistent tables
describe({ database: "persistent" })
sample({ table: "customers", database: "persistent" })

The database parameter is available on query, execute, load_data, load_file, load_files, watch_directory, describe, sample, chart, export, and set_table_metadata. The shorthand persist: true (sugar for database: "persistent") is available on load_data, load_file, load_files, and watch_directory. Pass any user-attached writable alias (created via attach_database) to target a custom database.

(query_data and query_file are one-shot tools that materialize the inline data into their own temp table and query it — they do not accept a database parameter because the data isn't in a persisted database to begin with.)

2. Fully-qualified SQL (for power users or complex multi-DB joins):

-- Read from persistent
SELECT * FROM "persistent"."public"."customers";

-- Write to persistent
CREATE TABLE "persistent"."public"."revenue_2026" AS
  SELECT region, SUM(amount) FROM scratch_orders GROUP BY region;

Per-database _table_catalog: every writable database — persistent and any user-attached writable file — gets its own _table_catalog lazily seeded on first ingest. MCP-managed metadata (load tool, params, timestamps, prose fields set via set_table_metadata) lives alongside the data file, so opening a .hyper file later as a primary workspace finds the catalog ready. If you want a pristine .hyper file for export with no MCP bookkeeping, run DROP TABLE "<alias>"."public"."_table_catalog" once and subsequent sessions opening that file will leave it dropped.

Detach safety: detach_database rejects with InvalidArgument if any active watcher targets the alias — call unwatch_directory first. This prevents the watcher's pool from silently writing into a now-detached file (or worse, the wrong file if the alias is later re-attached to a different path).

Daemon management

The daemon is normally invisible — it auto-spawns on first use and stays resident. For diagnostics:

hyperdb-mcp daemon status   # Show running daemon (PID, endpoint, started_at, version)
hyperdb-mcp daemon stop     # Gracefully shut down the daemon
hyperdb-mcp daemon          # Run as a daemon explicitly (rarely needed)

status and stop locate the running daemon automatically (reading daemon.json, then scanning the port range), so they work even if the daemon scanned onto a non-default port. Pass --port <PORT> to target a specific port explicitly.

State files live at ~/.hyperdb/ by default (override with HYPERDB_STATE_DIR).

Port discovery. The daemon binds a TCP health/lock port — by default it scans upward from 7485 (16 ports) and uses the first free one; set HYPERDB_DAEMON_PORT to pin an exact port (no scan). The health port doubles as a single-instance lock and an identity check: clients send PING and require a PONG hyperdb-mcp <version> reply before trusting a daemon, so an unrelated process occupying the port is skipped rather than mistaken for the daemon.

Staying resident. By default the daemon never idle-shuts-down — keeping hyperd warm means the next tool call connects immediately instead of triggering a "restarting, please retry" round-trip. To opt into auto-shutdown (e.g. on CI), pass --idle-timeout <SECS> or set HYPERDB_DAEMON_IDLE_TIMEOUT.

Recovery from hyperd crashes

The daemon polls hyperd every 5 seconds. If the process has exited (crashed, OOM, killed), the daemon spawns a replacement, atomically updates ~/.hyperdb/daemon.json with the new endpoint, and continues serving clients. Clients see one failed tool call (the request that was in flight when hyperd died); the next tool call transparently reconnects to the new hyperd via the same recovery path used for normal connection drops.

If a client itself notices hyperd is unreachable before the next polling tick, it sends a fast-path REPORT_HYPERD_ERROR signal to the daemon so the restart kicks off without waiting for the timer.

If hyperd repeatedly fails to start (3 attempts within 60 seconds — e.g., misconfigured HYPERD_PATH, port exhaustion, broken binary), the daemon shuts itself down and removes the discovery file. The next MCP client to start up will then spawn a fresh daemon, surfacing any persistent failure clearly to the user rather than spinning silently.

Known limitation: if hyperd hangs (alive at the OS level but unresponsive to queries), the daemon's polling can't detect it and your tool call may stall indefinitely. The recovery path is hyperdb-mcp daemon stop followed by reconnecting from your MCP client.

Other behavioral flags

Flag Behavior
--read-only Disables execute, load_data, load_file, watch_directory, save_query, delete_query, and Hyper-format export. See Read-Only Mode.

MCP Tools

One-Shot Tools

query_data

Ingest inline data and run a SQL query in a single call.

query_data(data: '[{"region":"West","revenue":1200},...]', sql: 'SELECT region, SUM(revenue) FROM data GROUP BY region')
Parameter Type Required Description
data string yes JSON array of objects, or CSV text
sql string yes SQL query to run against the data
format string no "json" or "csv" — auto-detected if omitted
table_name string no Table name for use in SQL — defaults to "data"
schema object no Partial column-name → type map (see Schema Overrides)

query_file

Ingest a file and run a SQL query in a single call. Streams from disk — handles files of any size.

query_file(path: '/tmp/sales.parquet', sql: 'SELECT TOP 10 * FROM sales ORDER BY amount DESC')
Parameter Type Required Description
path string yes Path to CSV / JSON / JSONL / Parquet / Arrow IPC file
sql string yes SQL query to run
table_name string no Table name — defaults to filename stem
schema object no Partial column-name → type map (see Schema Overrides)

Workspace Tools

load_data

Load inline data into a named workspace table.

load_data(table: 'customers', data: '[{"id":1,"name":"Alice"},...]')
Parameter Type Required Description
table string yes Table name
data string yes JSON array of objects, or CSV text
format string no "json" or "csv" — auto-detected
mode string no "replace" (default) or "append"
schema object no Partial column-name → type map (see Schema Overrides)

load_file

Load a file into a named workspace table.

load_file(table: 'orders', path: '/tmp/orders.csv')
Parameter Type Required Description
table string yes Table name
path string yes Path to CSV / JSON / JSONL / Parquet / Arrow IPC file
mode string no "replace" (default) or "append"
schema object no Partial column-name → type map (see Schema Overrides)

When you're unsure of the right types — or recovering from a previous SCHEMA_MISMATCH — call inspect_file first. It reports the exact schema load_file would use plus per-column min / max / null_count so you can build a minimal, correct override in one shot.

load_iceberg

Load an Apache Iceberg table into a named workspace table. Pass the absolute path to the Iceberg table root (the directory containing metadata/ and data/); hyperd's native Iceberg reader derives the schema and resolves the snapshot.

load_iceberg(table: 'sales', path: '/lake/warehouse/db/sales')
Parameter Type Required Description
table string yes Target Hyper table name
path string yes Absolute path to the Iceberg table root directory
mode string no "replace" (default) or "append"
metadata_filename string no Pin a specific snapshot, e.g. "v2.metadata.json". Omit for latest.
version_as_of integer no Pin a snapshot by version number

Schema overrides are not accepted — hyperd derives the schema from the Iceberg table metadata.

query

Run a read-only SQL query against the workspace. Accepts SELECT, WITH, EXPLAIN, SHOW, VALUES. For DDL/DML use execute.

query(sql: 'SELECT c.name, SUM(o.amount) FROM orders o JOIN customers c ON o.customer_id = c.id GROUP BY c.name')

execute

Execute one or more mutating SQL statements as an atomic batch: CREATE TABLE, INSERT, UPDATE, DELETE, DROP TABLE, ALTER, COPY, etc. sql is an array of statements; multi-element batches run inside a transaction (all commit or all roll back). Single-element batches auto-commit, same as a one-off statement. Returns the per-statement affected row counts plus a total. Disabled in read-only mode.

// Single statement (auto-commit)
execute(sql: ['CREATE TABLE archived_orders AS SELECT * FROM orders WHERE year < 2024'])

// Atomic upsert — both run or neither runs
execute(sql: [
  "UPDATE settings SET value = 'dark' WHERE key = 'theme'",
  "INSERT INTO settings (key, value) SELECT 'theme', 'dark' \
     WHERE NOT EXISTS (SELECT 1 FROM settings WHERE key = 'theme')"
])

Validation rules enforced before any SQL hits the server:

  • Array must be non-empty; no element may be empty / whitespace-only / comment-only.
  • No element may be read-only — use query for SELECT/WITH/EXPLAIN.
  • DDL and DML cannot be mixed in one batch (Hyper aborts mixed transactions with SQLSTATE 0A000).
  • Multi-element all-DDL batches are rejected because Hyper auto-commits CREATE/DROP/ALTER even inside a transaction; issue each DDL in its own execute call.
  • Explicit transaction-control statements (BEGIN / COMMIT / ROLLBACK / SAVEPOINT) in batch elements are rejected — the tool manages the transaction for you, and a user-issued COMMIT mid-batch would defeat atomicity.

describe

List all workspace tables with their schemas, column types, and row counts.

sample

Return the schema, total row count, and first N rows of a table in a single call.

sample(table: 'orders', n: 10)
Parameter Type Required Description
table string yes Table name
n int no Rows to return (default: 5, clamped to 1..=100)

Diagnostics

inspect_file

Dry-run schema inference on a CSV, Parquet, or Arrow IPC file without ingesting it. Returns the exact schema load_file / query_file would use (including the full-file numeric widening pass) plus per-column min, max, null_count, and sample_values. Nothing is written to Hyper and hyperd is not even started.

Use it before load_file whenever you are unsure about types, or after a SCHEMA_MISMATCH failure to pick the right widening. The LLM can feed the reported type + min / max directly into a partial schema override on the subsequent load_file call.

inspect_file(path: '/tmp/owid-population.csv')
Parameter Type Required Description
path string yes Path to CSV / JSON / JSONL / Parquet / Arrow IPC file
sample_rows int no Sample values / rows per column (default 5, clamped 1..=50)

Response shape:

{
  "file_format": "csv",
  "row_count": 63000,
  "file_size_bytes": 4831204,
  "columns": [
    { "name": "Entity",     "type": "TEXT",   "nullable": true, "null_count": 0,   "sample_values": ["Afghanistan", ...] },
    { "name": "Year",       "type": "INT",    "nullable": true, "null_count": 0,   "min": 1800, "max": 2023, "sample_values": ["1800", ...] },
    { "name": "Population", "type": "BIGINT", "nullable": true, "null_count": 12,  "min": 500, "max": 8002572256, "sample_values": ["4000000", ...] }
  ],
  "sample_rows": [ { "Entity": "Afghanistan", "Year": "1800", "Population": "2805829" } ]
}

sample_values and sample_rows are always strings, regardless of the inferred column type — they report what the file contains on disk, before any type coercion, so the LLM can compare the raw text against min / max when building a schema override. Use type (and min / max) for the typed view; use sample_values for the raw view.

Saved Queries

Register a named read-only SQL query once; read its live result as many times as you like via a resource URI. Useful for dashboard-style recurring views and for giving LLMs a stable "bookmark" set of key queries that resources/list advertises up front.

Each saved query produces two resources:

  • hyper://queries/{name}/definition — the stored SQL plus metadata (description, created_at) as JSON.
  • hyper://queries/{name}/result — re-runs the SQL on every read and returns { name, result: [...], stats: {...} }.

Persistence: saved queries land in the persistent attachment's _hyperdb_saved_queries meta-table ("persistent"."public"."_hyperdb_saved_queries") and survive server restarts. In --ephemeral-only sessions they live only for the lifetime of the server process.

save_query

save_query(name: 'top_5_customers', sql: 'SELECT customer, SUM(amount) AS total FROM orders GROUP BY customer ORDER BY total DESC LIMIT 5', description: 'Biggest spenders this year')
Parameter Type Required Description
name string yes Unique identifier used as the URI path component
sql string yes Read-only SQL (SELECT / WITH / EXPLAIN / SHOW / VALUES)
description string no Human-friendly summary

Duplicate names are rejected with INVALID_ARGUMENT — use delete_query first if you intend to overwrite. Non-read-only SQL is rejected with SQL_ERROR. Disabled in read-only mode.

delete_query

delete_query(name: 'top_5_customers')
Parameter Type Required Description
name string yes Name of the saved query to remove

Returns { "deleted": true } when the query existed, { "deleted": false } when it did not (no error on unknown names). Disabled in read-only mode.

Export Tools

export

Write query results or a table to a file.

export(table: 'orders', path: '~/Desktop/orders.parquet', format: 'parquet')
export(sql: 'SELECT ...', path: '~/Desktop/analysis.hyper', format: 'hyper')
Parameter Type Required Description
sql string no Query to export (if omitted, exports whole table)
table string no Table name (used if sql omitted)
path string yes Output file path
format string yes "csv", "parquet", "iceberg", "arrow_ipc", or "hyper"

The "hyper" format produces a .hyper file that opens directly in Tableau Desktop.

Visualization

chart

Render a chart from a SQL query and return it inline as an image.

chart(sql: 'SELECT product, SUM(revenue) as total FROM sales GROUP BY product', chart_type: 'bar', x: 'product', y: 'total', title: 'Revenue by Product')
Parameter Type Required Description
sql string yes Read-only SQL query returning the data to plot
chart_type string yes bar, line, scatter, or histogram
x string yes* X-axis column (for histogram, the value column)
y string yes* Y-axis column (not required for histogram)
series string no Grouping column for multi-series plots
title string no Chart title
format string no png (default) or svg
width int no Pixels (default 800, clamped 200..4096)
height int no Pixels (default 480, clamped 150..4096)
bins int no Histogram bins (default 20, clamped 1..500)

Returns an ImageContent (base64 PNG or SVG) plus a stats JSON block.

Incremental Ingest

watch_directory / unwatch_directory

Monitor a directory for data files and auto-append them to a target table.

watch_directory(path: '/tmp/inbox', table: 'events')
unwatch_directory(path: '/tmp/inbox')

Producer protocol (.ready sentinel):

  1. Write data file (e.g. foo.csv) and close it.
  2. Create a zero-byte companion foo.csv.ready — this is the atomic signal.
  3. Poll for the absence of foo.csv.ready to confirm the watcher is done.

On success, both files are deleted. On failure, both are moved to failed/ with a .error JSON file.

Key properties:

  • One directory, one table, append mode — files must match the target schema.
  • Initial sweep — pre-existing .ready files are processed immediately.
  • Read-only mode — watch_directory is blocked; unwatch_directory is always allowed.
  • Cleanup — dropping the server or calling unwatch_directory terminates the background thread.

Utility Tools

status

Returns plugin health, workspace mode, table count, total rows, disk usage, read-only flag, and active directory watchers with per-watcher stats.

MCP Resources

The server exposes workspace state as MCP Resources, discoverable via resources/list. Each resource advertises its own MIME type so clients can route it appropriately (LLM context vs. file download vs. chart).

URI MIME Content
hyper://workspace application/json Workspace mode, table count, total rows, disk usage
hyper://tables application/json Full list of tables with schemas and row counts
hyper://readme text/markdown Workspace overview as markdown: table catalog, related resources per table, and tool hints for a cold-started LLM
hyper://tables/{name}/schema application/json Columns, types, nullability, and row count for one table
hyper://tables/{name}/sample application/json First 5 rows of a table as JSON, with schema
hyper://tables/{name}/csv-sample text/csv First 20 rows of a table as CSV, header-first
hyper://queries/{name}/definition application/json Stored SQL + metadata for a saved query
hyper://queries/{name}/result application/json Live result of a saved query — re-runs on every read

Resource templates (discoverable via resources/templates/list):

  • hyper://tables/{name}/schema
  • hyper://tables/{name}/sample
  • hyper://tables/{name}/csv-sample
  • hyper://queries/{name}/definition
  • hyper://queries/{name}/result

The internal _hyperdb_saved_queries meta-table used to persist saved queries is deliberately hidden from resources/list and hyper://tables — callers see only user-visible data tables.

Resource-update notifications

HyperDB advertises both the resources.subscribe and resources.listChanged capabilities in its initialize response. Clients can subscribe to any hyper://... URI via resources/subscribe and will then receive notifications/resources/updated messages whenever the server detects a change, without polling.

The server fires targeted updates for the URIs affected by each kind of mutation:

Trigger Updated URIs resources/list_changed?
load_data / load_file (replace mode) hyper://workspace, hyper://tables, hyper://readme, per-table schema + sample + csv-sample Yes
load_data / load_file (append mode) Same per-table + summary URIs No ¹
watch_directory ingest of a .ready pair Same per-table + summary URIs No ¹
execute (INSERT / UPDATE / DELETE) Workspace summary URIs No
execute (CREATE / DROP / ALTER / TRUNCATE / RENAME) Workspace summary URIs Yes
save_query (none per-URI) Yes — two new hyper://queries/{name}/... resources
delete_query hyper://queries/{name}/definition, hyper://queries/{name}/result Yes — two resources disappeared

¹ Append-mode ingest (both load_* and the watcher) auto-creates the target table when it doesn't exist, but does not fire list_changed for that creation. Clients that need to discover watcher-created tables should re-read hyper://tables after subscribing, or use the per-table updated notification as a trigger to refresh their list. Tracked in DEVELOPMENT.md as tech debt.

Notifications are fire-and-forget — send failures (typically due to a client disconnect) are logged at the debug level and the registry prunes dead peers lazily. This keeps mutation paths fast and free of back-pressure concerns.

All JSON-typed resources return a pretty-printed object; Markdown and CSV resources are returned verbatim.

MCP Prompts

Four guided analytical workflows registered as MCP Prompts.

Prompt Arguments What it does
analyze-table table Schema walkthrough, column statistics, data quality flags
compare-tables table_a, table_b Schema alignment, JOIN key suggestions, analytical opportunities
data-quality table Systematic NULL / duplicate / cardinality / outlier checks
suggest-queries table, goal? 5 analytical SQL queries with explanations, optionally goal-guided

Read-Only Mode

hyperdb-mcp --persistent-db ~/analytics.hyper --read-only
  • Allowed: query, query_data, query_file, describe, sample, inspect_file, status, export
  • Blocked: execute, load_data, load_file, watch_directory, save_query, delete_query — return READ_ONLY_VIOLATION
  • Resources, prompts, and resource subscriptions work normally — read-only clients can still subscribe to hyper://... URIs and receive notifications when other (non-read-only) connections mutate state

The query tool also enforces read-only at the SQL level — only SELECT/WITH/EXPLAIN/SHOW/VALUES are accepted.

Data Flow Patterns

  • Small data (LLM relay): For <10K rows. The LLM gets data from another plugin and passes it inline via query_data.
  • Large data (file intermediary): For thousands to billions of rows. Source plugin exports to a file, the LLM calls query_file. Data never enters the LLM context — constant memory regardless of file size.

Schema Inference

Three tiers, chosen automatically based on the data source:

Tier Source How
Exact Arrow IPC, Parquet Schema read from file metadata. Types preserved exactly.
Structural JSON All objects scanned. Per-column type widening: Int → BigInt → Double. Mixed types → TEXT.
Heuristic CSV Header row for names, first 1,000 rows sampled for types. A second full-file streaming pass then widens numeric columns if needed (INT → BIGINT → NUMERIC(38,0); INT/BIGINT → DOUBLE PRECISION if any later row contains a decimal).

JSON file shapes. load_file and query_file accept two JSON representations and auto-detect between them from the first non-whitespace byte: a top-level JSON array of objects (e.g. [{...}, {...}]) or newline-delimited JSON (JSONL / NDJSON — one JSON object per line, the format hyperd's own logs use). Blank lines are tolerated. Malformed JSONL surfaces a SCHEMA_MISMATCH error naming the offending line number.

Content sniffing for unknown extensions. Files with extensions the dispatcher doesn't recognize (.log, .txt, no extension at all) are classified by peeking at the first non-whitespace byte: [ or { routes to JSON, anything else to CSV. This means hyperd's raw .log files load through load_file directly, no rename or preprocessing required. Binary formats (.parquet, .arrow, .ipc, .feather, .pq) always win by extension since they're not text-sniffable. inspect_file uses the exact same dispatcher so its report can never disagree with what load_file would do.

CSV NULL handling. Unquoted empty cells (,,) load as SQL NULL — matching PostgreSQL's CSV convention and inspect_file's null_count diagnostics. Quoted empty strings (,"",) load as the literal empty string. This means downstream WHERE col IS NULL works directly without a defensive OR col = '' clause.

The full-file CSV widening pass specifically protects against the "big value hidden at the end of the file" failure mode — e.g. an aggregate row whose population is ~8 billion tucked in after 60 000 country-sized rows. Without it, the first-pass sample would pick INT and the COPY would fail with SCHEMA_MISMATCH / SQLSTATE 22003 mid-ingest.

For implementation details (widening rules, type mapping tables), see the module docs in src/schema.rs and src/ingest_arrow.rs.

Schema Overrides

Every data-in tool (query_data, query_file, load_data, load_file) accepts an optional schema parameter: a partial map from column name to Hyper SQL type.

{ "schema": { "population": "BIGINT", "order_date": "DATE" } }

Semantics:

  • Keys are matched to columns by name (case-sensitive). Column order in the JSON object does not need to match the file — the inferred order from the file is preserved.
  • Columns not listed in the override keep their inferred type. You only specify the columns you want to correct.
  • Unknown column names and unknown type strings are rejected up front with a SCHEMA_MISMATCH error that lists the real column names, so the LLM can self-correct without another round-trip.
  • Supported type strings: INT, BIGINT, NUMERIC(p,s) (e.g. NUMERIC(38,0) or NUMERIC(12,2)), DOUBLE PRECISION, TEXT, BOOL, DATE, TIMESTAMP.

Recommended workflow for unfamiliar data:

  1. Call inspect_file → read the reported type + min / max per column.
  2. For any column whose max exceeds its inferred type's range, or where you want stricter parsing than CSV heuristics give, build a partial override.
  3. Pass it to load_file / query_file.

SQL Dialect

Hyper uses the Salesforce Data Cloud SQL dialect (PostgreSQL-compatible with extensions). Supports SELECT, JOINs, subqueries, CTEs, window functions, aggregations, DDL, DML, and COPY FROM.

Upserts (INSERT or UPDATE)

Hyper does not support ON CONFLICT or INSERT ... ON DUPLICATE KEY. Use the execute tool's atomic batch shape instead:

execute(sql: [
  "UPDATE settings SET value = 'dark' WHERE key = 'theme'",
  "INSERT INTO settings (key, value) SELECT 'theme', 'dark' \
     WHERE NOT EXISTS (SELECT 1 FROM settings WHERE key = 'theme')"
])

Both statements run inside a single Hyper transaction — they commit together or both roll back. No race window between them.

Tip: For file-based upserts (merging updated data from a CSV/JSON file into an existing table), use load_file with mode: "merge" and a merge_key instead of writing manual SQL — it handles the UPDATE/INSERT logic automatically and also auto-adds new columns.

Transactions

The Hyper Rust API supports BEGIN / COMMIT / ROLLBACK plus an RAII Transaction guard (see docs/TRANSACTIONS.md). The MCP execute tool surfaces this as the sql array shape: pass multiple statements and they run atomically.

Hyper-specific limits worth remembering when batching:

  • DDL after DML in the same transaction is rejected with SQLSTATE 0A000. The execute tool catches this up front — mixing CREATE/DROP/ALTER with INSERT/UPDATE/DELETE in one batch is rejected with an actionable error.
  • DDL is auto-committed even inside a transaction. execute rejects multi-element all-DDL batches because the "atomic" promise can't be honored — issue each DDL call as its own one-element array.
  • After any error inside a transaction, the connection enters aborted state and only ROLLBACK is accepted next. The execute tool handles this for you — on any per-statement failure the wrapper issues ROLLBACK before surfacing the error.

Full reference: Data Cloud SQL Reference

CLI Reference

hyperdb-mcp [OPTIONS] [COMMAND]

Commands:
  daemon                  Run as a background daemon managing a shared hyperd process

Options:
  --persistent-db <PATH>  Path to the persistent .hyper file. Defaults to the platform
                          data dir (~/Library/Application Support/hyperdb/workspace.hyper
                          on macOS, ~/.local/share/hyperdb/workspace.hyper on Linux,
                          %APPDATA%\hyperdb\workspace.hyper on Windows). Override via
                          the HYPERDB_PERSISTENT_DB env var.
  --ephemeral-only        Skip the persistent attachment entirely. Disables save_query
                          persistence (queries fall back to session storage).
  --read-only             Disable mutating tools (execute, load_data, load_file,
                          save_query, delete_query, watch_directory)
  --no-daemon             Disable the shared daemon and spawn a private hyperd

Deprecated:
  --workspace <PATH>      Old name for --persistent-db. Still accepted, emits a
                          stderr warning, and will be removed in a future release.

Daemon subcommand:
  hyperdb-mcp daemon                          Start the daemon (usually auto-spawned)
  hyperdb-mcp daemon stop                     Gracefully stop the running daemon
  hyperdb-mcp daemon status                   Show running daemon info
  hyperdb-mcp daemon --port <PORT>            Pin the health/lock port. When omitted,
                                              scans upward from 7485 for a free port.
  hyperdb-mcp daemon --idle-timeout <SECS>    Opt into idle shutdown after SECS idle.
                                              When omitted, the daemon stays resident.

Environment:
  HYPERD_PATH                  Path to hyperd binary (auto-detected if on PATH)
  HYPERDB_PERSISTENT_DB        Override the default persistent-db path
  HYPERDB_STATE_DIR            Override daemon state directory (default ~/.hyperdb/)
  HYPERDB_DAEMON_PORT          Pin daemon health/lock port (default: scan from 7485)
  HYPERDB_DAEMON_IDLE_TIMEOUT  Opt into idle shutdown (seconds); default: stay resident

Error Handling

Errors include a machine-readable code and a suggestion:

Code When Recovery
HYPERD_NOT_FOUND hyperd not found Set HYPERD_PATH or install Hyper
FILE_NOT_FOUND File path doesn't exist Verify the path
UNSUPPORTED_FORMAT Unrecognized file type Specify format explicitly
SCHEMA_MISMATCH Data doesn't match inferred types, numeric overflow (SQLSTATE 22003), or invalid text for target type (SQLSTATE 22P02) Call inspect_file then retry with a partial schema override (e.g. {"population":"BIGINT"} or {"id":"TEXT"})
SQL_ERROR Invalid SQL Fix the query
TABLE_NOT_FOUND Table doesn't exist Use describe to list tables
READ_ONLY_VIOLATION Mutating op in read-only mode Use query_* / inspect_file, or restart without --read-only
CONNECTION_LOST hyperd crashed or wire protocol desynchronized Retry — the server tears down the engine and reconnects on the next call

Server-returned errors include a machine-readable code, a message, and a suggestion with concrete retry guidance. The SCHEMA_MISMATCH suggestion for an overflow names the workflow directly: "call inspect_file, then retry with a partial schema override", so the LLM does not need to infer the recovery steps from the SQLSTATE alone.

Troubleshooting

Tools not discovered by the client — Verify the initialize response advertises "capabilities": {"tools": {}}. Pipe a raw initialize JSON-RPC request to the binary to check.

Server registered but tools not callable (Claude Code) — Add "mcp__HyperDB__*" to the permissions.allow array in ~/.claude/settings.json.

hyperd not found — Set HYPERD_PATH in the MCP server's env config, or place hyperd on your PATH.

Related Documentation

  • Main README — Getting started with the Hyper API
  • hyperdb-api — Core Rust API (sync/async connections, inserter, query)
  • DEVELOPMENT.md — Internal architecture, design decisions, contributor guide
  • ROADMAP.md — Forward-looking design sketches for features that aren't built yet
  • Design Spec — Full design document