This server provides semantic context about a data warehouse, enabling LLMs to find, understand, trace, and assess the health of tables beyond just schema information.
search_tables: Find tables by describing what you want in plain English. Returns ranked candidates with purpose, owner, and status (active/deprecated/experimental), including deprecated tables so you can avoid them.
describe_table: Get the full meaning of a table: purpose, grain (what a row represents), owner, materialization, row count, last updated, and column details (type, semantic units like "USD cents (net)" vs. "count", null percentage, description, primary key flag). This prevents guessing about gross/net, cents/dollars, etc.
trace_lineage: Trace upstream sources and downstream dependencies for a table or column, up to 5 hops deep. Provides column-level lineage when possible via SQL parsing; any fallback to model-level is noted in unresolved_notes.
check_health: Assess a table's reliability with a real-time verdict (healthy, stale, failing, deprecated), freshness in hours, row count and 7-day delta, per-column null rates, dbt test results, and a one-line summary.
It leverages both automatically derived metadata (lineage, freshness, row counts) and human-authored dbt inputs (units, deprecation status) to help LLMs avoid guesswork.
Exposes semantic metadata from dbt projects, including table purpose, ownership, column descriptions, units, deprecation status, and column-level lineage derived from dbt artifacts and compiled SQL.
Enables live health and freshness checks on DuckDB tables, including row counts, 7-day row deltas, null rates, and dbt test results, so agents can assess whether a table is safe to rely on.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@semantic-context-mcp Which table holds daily net revenue?"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
An MCP server that gives an LLM the meaning of a data warehouse, not just its schema.
A model with schema access sees a table called tbl_rev_fnl_v2 and a column called amt. It has
no way to know whether that's gross or net, cents or dollars, or a table nobody has touched in
40 days — so it guesses, and the guess runs and returns a plausible, wrong number. This server
closes that gap by exposing four tools backed by real derived metadata: ownership, freshness,
column units, deprecation status, and column-level lineage.
Stack: Python · Model Context Protocol · dbt + DuckDB · SQLite (FTS5) · sentence-transformers · sqlglot · Anthropic API (tool use / eval harness) · pytest
Tool | Answers |
| "Which table holds daily net revenue?" |
| "What does one row mean, and what unit is this column in?" |
| "Where did this number come from, and what breaks if I change it?" |
| "Is this table safe to rely on right now?" |
time — see "Setup" below. Without it, search_tables still works, just as keyword-only BM25.
The credibility of a project like this rests entirely on whether the "meaning" it reports is real or hand-typed for the demo. It's mostly real:
Signal | Source | Human-authored? |
Table purpose, grain, materialization |
| No |
Test pass/fail | dbt build results | No |
Model-level lineage |
| No |
Column-level lineage |
| No |
Table search (keyword) | SQLite FTS5 / BM25 over name + description + columns | No |
Table search (semantic) | Cosine similarity over MiniLM embeddings | No |
Row counts, 7-day row deltas | Live DuckDB queries at ingestion time | No |
Freshness (hours since latest data) | Recomputed live on every call | No |
Null rates per column | Live DuckDB queries at ingestion time | No |
Ownership | dbt | Partly |
Column descriptions and units | dbt | Yes — the same input any dbt project already has |
Deprecation status + successor | dbt | Yes |
Only two rows are human-authored, and both are ordinary dbt metadata any real project already maintains — this isn't reading from a hand-curated demo file.
Column-level lineage resolves 35 of 36 columns (97%) in the fixture warehouse. The one exception
sums a column with no table-qualifying prefix across a JOIN — genuinely ambiguous SQL that sqlglot
correctly declines to guess on, falling back to model-level lineage instead. trace_lineage always
names exactly where a fallback happened, in unresolved_notes — never silently.
The demo warehouse is dbt-labs/jaffle_shop_duckdb
with four added "decoy" revenue models, since the stock fixture is too clean to show any ambiguity.
The decoys are synthetic and labelled as such in fixture/ — the pipeline that derives
metadata from them is not.
Related MCP server: Semantic BI MCP
Requires Python 3.12+ and uv. No GPU needed.
uv syncThis installs only what the server itself needs and gives you a fully working server:
describe_table, trace_lineage, and check_health are fully live, and search_tables works in
keyword (BM25) mode. Heavier, optional pieces are separate extras so the base install stays fast:
uv sync --extra search # sentence-transformers + numpy -- semantic ranking for search_tables
uv sync --extra ingest # sqlglot + duckdb -- only needed to (re)build catalog.db
uv sync --extra dbt # dbt-core + dbt-duckdb -- only needed to build the fixture warehouse
uv sync --extra eval # anthropic + pyyaml -- only needed to run the eval harness
uv run <anything>re-syncs the environment to match whatever--extraflags that invocation passes — running a bareuv run python ...after syncing extras in can silently drop them again. If a tool that was working starts raisingModuleNotFoundError, re-run theuv sync --extra ...line below rather than debugging the code.
Needed once before the server has data to serve — without it, every tool call fails with a clear error telling you to run this.
uv sync --extra dbt --extra ingest --extra search
uv run python fixture/build_warehouse.py # seed -> dbt build -> dbt docs generate
uv run python -m semantic_context_mcp.ingest # manifest + catalog + run_results -> catalog.dbRe-running this sequence is always safe — both scripts fully replace their outputs rather than
merging into them, and build_warehouse.py regenerates seed dates so the "stale" decoy table
always looks stale relative to today. dbt's very first invocation in a fresh venv can sit silent
for 20-30 seconds while it compiles bytecode for its own dependency tree before printing anything —
build_warehouse.py prints a heads-up before that step so it doesn't look hung; subsequent runs are
fast.
uv run semantic-context-mcpThe server speaks MCP over stdio and sits silently waiting for a client — that's correct behavior, not a hang. To exercise it directly without a client:
uv run python scripts/smoke_stdio.pyThat script spawns the real server over the real transport and drives every tool through both happy-path and error-path calls, printing each result.
uv run python -c "
from semantic_context_mcp.core import search, describe, health
for m in search.run('daily net revenue'):
print(m.name, '-', m.purpose)
print()
print(describe.run('fct_revenue_daily').columns)
print()
print(health.run('tbl_rev_fnl_v2').summary)
"Add to claude_desktop_config.json
(%APPDATA%\Claude\claude_desktop_config.json on Windows,
~/Library/Application Support/Claude/claude_desktop_config.json on macOS):
{
"mcpServers": {
"semantic-context": {
"command": "uv",
"args": ["--directory", "/absolute/path/to/semantic-context-mcp", "run", "semantic-context-mcp"]
}
}
}Restart Claude Desktop, then ask "What's our daily revenue table?"
uv run pytestThe suite covers synthetic unit tests (fast, no fixture needed, always run) and real-data
integration tests that build the actual dbt/DuckDB warehouse and ingest it into catalog.db — the
latter are what caught every real bug during development (a dbt artifact-ordering issue, a DuckDB
NULL-vs-zero aggregation bug, a naive/aware datetime crash, a missing test hiding a staleness
signal, and a false "lineage unresolved" report on a fully-resolved column). Two additional
scripts exercise the protocol layer directly rather than through pytest:
uv run python scripts/smoke_stdio.py # real server, real MCP stdio transport
uv run python -m semantic_context_mcp.eval.mcp_path_smoke # proves the eval harness matches the real servereval/ measures whether having this server's tools available actually changes whether Claude picks
the right table and states the right caveats, compared to schema access alone. Three arms, the same
25 questions, the same model and system prompt — only the tool set differs:
Arm | Tools |
| table + column names and SQL types only — no metadata (the realistic baseline) |
| schema-only tools, plus the four semantic tools |
| the four semantic tools alone |
uv sync --extra eval
export ANTHROPIC_API_KEY=... # or `ant auth login`
uv run python -m semantic_context_mcp.eval # all 3 arms, 25 questions x 3 runs
uv run python -m semantic_context_mcp.eval --questions 5 --runs 1 # a cheap smoke run
uv run python -m semantic_context_mcp.eval --dry-run # exercises the harness, no API callsEach question/run/arm writes a transcript to eval/runs/<timestamp>/, so any surprising result is
checkable rather than trusted. Grading (eval/grading.py) is deliberately simple keyword/table-name
matching over the model's final answer — not an LLM judge — and is fully unit-tested without any
network access.
fixture/
jaffle_shop_duckdb/ Cloned dbt project + models/decoys/*.sql (4 synthetic ambiguous tables)
pristine_seeds/ Untouched original seed CSVs -- prepare_seeds.py's source of truth
prepare_seeds.py Shifts seed dates to end "yesterday", idempotently
build_warehouse.py seed -> dbt build -> dbt docs generate
src/semantic_context_mcp/
models.py Pydantic contract for all four tool returns
errors.py Errors written to be read by the model
server.py MCP adapter -- no logic
core/ The actual logic; also what eval/tools.py calls
ingest/ dbt artifacts + git + SQLite -> catalog.db
eval/ Three-arm measurement harness (see "Evaluation" above)
scripts/
smoke_stdio.py Drives the real server over the real MCP transport
tests/ Contract, consistency, ingestion, eval, and stdout-purity guardsCheck whether a table is safe to rely on right now.
Call this before presenting any number derived from a table you have not already checked, and whenever a user mentions a table you did not choose yourself. Reports freshness, row-count movement, null rates, and dbt test results, plus a one-line verdict. A verdict of "deprecated" or "failing" means you should warn the user before using the table, not silently proceed.
Args: table: Table name, fully qualified or bare.
| Name | Required | Description | Default |
|---|---|---|---|
| table | Yes |
| Name | Required | Description |
|---|---|---|
| notes | No | |
| table | Yes | |
| tests | No | |
| summary | Yes | One line a human can act on |
| verdict | Yes | |
| row_count | No | |
| checked_at | No | |
| null_rates | No | Column name -> percentage of NULL values |
| freshness_hours | No | Hours since the most recent event in the table |
| row_count_delta_7d | No | |
| max_event_timestamp | No |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description carries the full burden. It discloses what the tool reports (freshness, row-count movement, null rates, dbt test results, one-line verdict) and how to act on the verdict, which goes beyond a vague 'checks health'. It does not explicitly state read-only behavior or side effects, but the nature of the tool implies a safe inspection. This is solid but not exhaustive.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is front-loaded with the main purpose, followed by clear usage guidance, a concise list of reported metrics, interpretation instructions, and the parameter description. Every sentence adds value and there is no redundancy or filler.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the tool has a single parameter and an output schema, the description provides ample context: what the tool does, when to use it, what it returns (including specific metrics and verdict), and how to interpret results. It is fully self-contained and leaves no critical gaps for an agent to operate correctly.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema coverage is 0%, but the description compensates with an Args section: 'table: Table name, fully qualified or bare.' This adds format guidance beyond the bare type of 'string' in the schema, giving the agent enough to construct a valid call.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description begins with a specific verb+resource: 'Check whether a table is safe to rely on right now.' This clearly distinguishes the tool from siblings like search_tables, describe_table, and trace_lineage, which address different concerns (discovery, structure, lineage).
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description gives explicit when-to-use guidance: 'Call this before presenting any number derived from a table you have not already checked, and whenever a user mentions a table you did not choose yourself.' It also explains how to interpret verdicts. However, it does not explicitly state when not to use the tool or name alternatives, so it falls slightly short of a 5.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Get the full meaning of one table: purpose, grain, owner, status, and columns.
Call this before using any table you have not already described in this
conversation. The column unit field is the part that matters most — it
distinguishes gross from net and cents from dollars, which the schema alone
cannot tell you. A unit of "unknown" means nobody documented it; say so
rather than guessing.
Args: table: Table name, fully qualified or bare (e.g. "main.fct_revenue_daily"). include_columns: Set False when you only need purpose, owner, and status.
| Name | Required | Description | Default |
|---|---|---|---|
| table | Yes | ||
| include_columns | No |
| Name | Required | Description |
|---|---|---|
| name | Yes | |
| grain | No | What one row represents, e.g. 'one row per customer per day' |
| owner | No | |
| status | No | |
| columns | No | |
| purpose | Yes | |
| row_count | No | |
| truncated | No | True when the column list was cut to protect the context window |
| last_updated | No | |
| superseded_by | No | |
| materialization | No | table, view, incremental, ... |
| truncation_hint | No | How to retrieve the omitted columns when truncated is True |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries the full burden. It discloses the interpretive behavior of the `unit` field, explaining that it distinguishes gross/net and cents/dollars, and instructs the agent to say 'unknown' rather than guess when the unit is undocumented. This adds valuable context beyond the schema, but it doesn't cover other potential behaviors like error handling or access requirements.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is front-loaded with the purpose, followed by usage guidance and a clearly structured Args section. Every sentence adds information, from the unit field warning to the parameter explanations, with no filler. It's concise yet complete.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
The description covers the essential context for using this tool: when to call it, what to expect (purpose, grain, owner, status, columns), and the key interpretation caveat about the unit field. Since an output schema is present, it doesn't need to enumerate return values. It's complete for the tool's complexity.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The schema has no descriptions for the parameters (0% coverage), but the description compensates fully with an Args section. For `table`, it provides an example of a fully qualified name; for `include_columns`, it explains when to set it to False. This gives the agent the needed semantic detail.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description starts with a clear, specific statement: 'Get the full meaning of one table: purpose, grain, owner, status, and columns.' This immediately identifies the verb, resource, and scope, distinguishing it from sibling tools like search_tables, which searches across tables, and trace_lineage, which follows relationships.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description provides a clear usage rule: 'Call this before using any table you have not already described in this conversation.' It also gives conditional guidance for include_columns, saying to set False when only purpose, owner, and status are needed. However, it doesn't explicitly mention alternative tools or describe when not to use it beyond the already-described condition, so it lacks the full 5-level criteria.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Find warehouse tables by describing what you want in plain English.
Call this before writing SQL whenever you are not certain which table holds the data — table names alone do not tell you whether a table is current, correct, or maintained. Returns candidates ranked by relevance, each with its owner and status. Deprecated tables are included and labelled, so you can recognize them in existing queries rather than assuming they are fine.
Args: query: What you are looking for, e.g. "daily net revenue" or "customer churn". limit: Maximum candidates to return (1-25).
| Name | Required | Description | Default |
|---|---|---|---|
| limit | No | ||
| query | Yes |
| Name | Required | Description |
|---|---|---|
| result | Yes |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description carries the full burden of behavioral disclosure. It goes beyond a simple 'search' by revealing that results include deprecated tables (labeled), are ranked by relevance, and include owner and status. This helps the agent understand the return value and potential pitfalls, exceeding the minimum expected for a read-only search tool.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is appropriately sized and well-structured: it opens with the purpose, then provides usage context, then describes return behavior, and finally lists arguments with examples. Every sentence adds value, with no redundancy or filler.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the tool has only two simple parameters and no annotations, the description covers all essentials: when to call it, what it returns, and how parameters behave. The existence of an output schema means the description need not detail the return format, but it still mentions key output traits. It is complete for the tool's complexity.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 0%, so the description must compensate. It fully explains both parameters: 'query' is described as 'What you are looking for, e.g. ...' and 'limit' is described as 'Maximum candidates to return (1-25).' This adds meaningful constraints and examples that are absent from the input schema.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description starts with a clear, specific verb phrase: 'Find warehouse tables by describing what you want in plain English.' It identifies the resource (warehouse tables) and the method (natural language query), distinguishing it from sibling tools like describe_table or trace_lineage. The purpose is immediately obvious.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description gives explicit guidance on when to use the tool: 'Call this before writing SQL whenever you are not certain which table holds the data.' It explains why table names alone are insufficient, but it does not explicitly name alternatives or state when not to use it. Clear context is provided, but exclusion/alternative guidance is missing.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Trace where a table or column's data comes from, and what depends on it.
Call this when you need to justify a number, when a value looks wrong and
you need to find where it was computed, or before suggesting a change to a
table so you can say what it would break. Pass column for column-level
lineage. Edges marked model_only mean the SQL could not be parsed down to
the column — the dependency is real, the precision is not.
Args: table: Table name, fully qualified or bare. column: Optional column to trace; omit for table-level lineage. direction: "up" for sources, "down" for dependents, "both" (default). depth: How many hops to follow (1-5).
| Name | Required | Description | Default |
|---|---|---|---|
| depth | No | ||
| table | Yes | ||
| column | No | ||
| direction | No | both |
| Name | Required | Description |
|---|---|---|
| depth | No | |
| table | Yes | |
| column | No | |
| upstream | No | |
| direction | No | |
| downstream | No | |
| unresolved_notes | No | Places where column-level resolution failed and why |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries the full burden. It discloses a key behavioral nuance: 'Edges marked model_only mean the SQL could not be parsed down to the column — the dependency is real, the precision is not.' This goes beyond the basic operation and warns about data-quality limitations.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is compact, front-loaded with the primary purpose, then uses a short usage-rule sentence and a tight Args list. No filler; every sentence and parameter comment serves a purpose.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Despite lack of annotations, the description covers purpose, usage triggers, parameter semantics, and a precision caveat. Since an output schema exists, the absence of return-format details is acceptable. The tool is fully specified for an agent to select and invoke correctly.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The schema has 0% description coverage, but the Args section explains every parameter: table as 'fully qualified or bare', column as optional for table-level lineage, direction with valid values, and depth with a hop range. This adds meaning beyond the bare schema names and defaults.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description opens with a specific verb and resource: 'Trace where a table or column's data comes from, and what depends on it.' This clearly defines lineage tracing and distinguishes it from sibling tools like search_tables, describe_table, and check_health.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description gives explicit trigger scenarios: 'when you need to justify a number, when a value looks wrong and you need to find where it was computed, or before suggesting a change to a table so you can say what it would break.' It does not name alternatives directly (e.g., use describe_table for schema-only questions), but the context is clear enough.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
v0.1.0check_healthdescribe_tablesearch_tablestrace_lineageScored across 4 tools
Each tool addresses a distinct aspect of table understanding: search finds tables by description, describe explains meaning and columns, trace_lineage maps data flow, and check_health assesses reliability. There is no overlap between these purposes, so an agent can cleanly select the right tool for the job.
All tool names follow a consistent verb_noun pattern with lowercase and underscores: search_tables, describe_table, trace_lineage, check_health. The verbs are specific and the nouns clearly indicate the target resource, making the API predictable.
Four tools is a tight, well-scoped set for a semantic context server focused on warehouse table understanding. Each tool covers a necessary capability without redundancy or bloat, fitting comfortably within the ideal 3-15 range.
The tool surface covers the complete workflow for exploring and validating tables: discover via search, inspect via describe, understand dependencies via lineage, and assess trustworthiness via health. There are no obvious missing operations that would force an agent to work around gaps.
MCP server for querying and analyzing data from ad platforms, analytics tools, and spreadsheets
MCP server for building and testing AI agents with multi-model experimentation and insights.
The BigQuery remote MCP server is a fully managed service that uses the Model Context Protocol to connect AI applications and LLMs to BigQuery data sources. It provides secure, standardized tools for AI agents to list datasets and tables, retrieve schemas, generate and execute SQL queries through natural language, and analyze data—enabling direct access to enterprise analytics data without requiring manual SQL coding.
The AI that understands raw data: Schema over your tables and databases, as MCP tools.
