# Bonnard vs Cube: MCP server features compared

> Bonnard and Cube MCP servers compared: what the agent knows about your data, how wrong answers get fixed, who can connect and pricing.

With Cube (cube.dev), the agent names a measure and Cube writes the warehouse SQL. With Bonnard (bonnard.ai), the agent reads each metric's SQL with your schema and writes the query itself.

## Five differences

### What the agent knows about your data

Bonnard gives its context to the AI client the person already uses. `explore_schema` searches table names, descriptions, column comments, enum labels, metrics and business terms, and returns up to 10 tables. For each table the agent gets the keys, row count and sample rows from the database, plus what your team adds: descriptions, joins, metrics, business terms and the values of columns you mark as searchable ([context](/features/context-engine)). A metric is its calculation in SQL, with instructions. The agent reads it and writes the query. Bonnard does not rewrite the query to apply it.

Cube gives outside AI clients its semantic model. `searchDataModel` searches views and their measures and dimensions, and returns the name, title, description and type of each ([Cube's MCP docs](https://docs.cube.dev/docs/integrations/mcp-server)). `runQuery` takes Cube SQL on those views, for example `MEASURE(revenue)`, and Cube writes the warehouse SQL from the measure's definition. The definition is enforced: the 30 MCP tools include none that runs SQL on warehouse tables. `runQuery` is marked read-only, and queries keep the user's permissions and row-level security.

Cube holds more context for its own agent. [Rules](https://docs.cube.dev/admin/ai/rules) are instructions in Markdown files. [Certified queries](https://docs.cube.dev/admin/ai/certified-queries) are approved SQL examples that the agent uses or adapts. The `meta.ai_context` field on a view or member is text that only the agent reads ([AI context](https://docs.cube.dev/docs/data-modeling/ai-context)). Claude or ChatGPT reaches that agent with the `chat` tool. Roles that can edit the model, Admin and Developer by default, can also read the model's source files with `readDataModelFile`.

### How wrong answers get found and fixed

Bonnard runs a loop of four steps on how agents use your data. Cube has records for the first step and evals for the last, and a person does the two steps between by hand.

[Image: The context loop in four steps: observe, diagnose, improve, verify.]

**Observe.** Bonnard records each tool call with its SQL, status and error, from every AI client your people use.

[Chats History](https://docs.cube.dev/admin/monitoring/chats-history) shows admins each conversation with Cube's agent, with a filter for those that came through MCP. [Query History](https://docs.cube.dev/admin/monitoring/query-history) lists queries with their source, errors and generated SQL, and its retention runs from 1 to 30 days by tier ([pricing docs](https://docs.cube.dev/admin/account-billing/pricing)). [Usage Analytics](https://docs.cube.dev/admin/monitoring/usage-analytics), on Premium and above, has dashboards for AI messages and token use.

**Diagnose.** Bonnard reads that record and finds what is wrong: definitions that differ between people, searches that found nothing, sessions that took many calls, repeated errors and tables that stopped updating. Where the data cannot settle a question, Bonnard asks your data team.

[Image: Bonnard's activity record over 30 days. It finds three definitions of net revenue across eight people, a search for gift cards that found nothing, sessions that took nine calls, an undefined column error and a table that stopped in May.]

On Cube, a data team reads those records by hand. Cube's docs list no feature that suggests context changes from agent activity, and no rating of answers by users.

**Improve.** Your data team's answer goes into a draft of the context.

A data team on Cube edits the model, rules and certified queries by hand. A person can certify a query from a chat. Cube's agent also stores memories from conversations and recalls them later ([memory](https://docs.cube.dev/admin/ai/memory-isolation)).

**Verify.** Evals run the same questions against the draft and the live version, and show which answers got better. A person then publishes the draft, and Bonnard keeps each version.

[Image: Three questions, each run five times against the live context and against the draft. Two get better and one shows no clear difference.]

[Evals](https://docs.cube.dev/admin/ai/evals) grade Cube's agent. Each question is a YAML entry with a certified query or SQL as the correct answer. The grader runs both queries and compares the rows, and an LLM judge can review the failures. Evals run on any branch, from the model IDE, the Cube CLI or the API. Running them in CI to gate a pull request is in preview, and Cube support turns it on.

Two warnings also reach the agent as it works. A result over the row limit tells the agent not to sum or average it. On PostgreSQL, a join that could count rows twice gets a warning.

Cube's evals grade Cube's agent. Bonnard's loop improves what Claude, ChatGPT or Cursor read before they write a query.

### What you need first

Bonnard starts from your database. Add it as a data source through a read-only credential, then give people access. The agent reads the table comments, keys and enum values that the database declares, and the context your team adds in Bonnard.

Cube starts from a semantic model in Cube Cloud: views with measures and dimensions, over one of 29 [data sources](https://docs.cube.dev/admin/connect-to-data/data-sources). An agent can only query what a view exposes. A Developer or Admin can have an agent add a missing measure on a development branch, then commit and merge it ([Cube's MCP docs](https://docs.cube.dev/docs/integrations/mcp-server)).

### Who can connect

A Bonnard workspace has unlimited users, your customers' users included. Each customer's query runs as that customer's tenant (the value that marks its data in your database). A customer with no tenant gets no data. Isolation uses the database: row-level security on PostgreSQL, Redshift and SQL Server, or one dataset or database for each tenant on BigQuery and Azure SQL Database ([tenancy](/docs/access/tenancy)).

On Cube, each person signs in through OAuth with their own Cube account, at the Viewer role or higher ([Cube's MCP docs](https://docs.cube.dev/docs/integrations/mcp-server)). A backend app can sign in with an API key, and an admin's key can act as any existing Cube user. Cube's MCP docs give no route for people without a Cube user. Ask Cube how its embedded products cover your customers' users.

### How you pay

Bonnard bills for each call, with no seats or minimums. Free includes 3,000 calls a month, then USD 0.95 for each 100 calls once an admin turns on extra usage ([pricing](/pricing)).

Cube's MCP server is on all plans. Free has developer seats and a limit of 1,000 requests a day. Starter is USD 40 for each developer a month. Premium is USD 80 for each developer a month, USD 40 for each explorer and USD 20 for each viewer, and the Explorer and Viewer roles start there ([Cube pricing](https://cube.dev/pricing)). Creating workbooks over MCP needs the Explorer role or higher. Compute costs extra: a dedicated deployment is USD 0.60 an hour on Starter and USD 1.20 on Premium. Cube's agent uses AI tokens. Each seat has a monthly grant, and use above it is billed on demand, up to a limit an admin sets ([AI tokens](https://docs.cube.dev/admin/account-billing/ai-tokens)).

## Side by side

Table: Bonnard and Cube MCP servers, verified 16 Nov 2026

| Feature | Bonnard | Cube |
| --- | --- | --- |
| Who answers | The AI client the person uses, with Bonnard's context | The AI client, with `searchDataModel` and `runQuery`, or Cube's agent through `chat` |
| Context the agent reads | Schema, descriptions, joins, metrics, business terms, searchable values | Views, measures and dimensions with descriptions. Rules, certified queries and `ai_context` for Cube's agent |
| Context through MCP | `explore_schema`, for any AI client | `searchDataModel`, for any AI client. Rules and certified queries through `chat` |
| Metric definitions | The agent reads the SQL and writes the query | Enforced. The agent names a measure and Cube writes the warehouse SQL |
| Observe | Records each tool call, from any AI client | Chats History, Query History and Usage Analytics |
| Diagnose | Finds what is wrong in the activity record | By hand |
| Improve | The data team's answer goes into a draft | By hand |
| Verify | Evals: the draft against the live version | Evals against a correct query, on any branch, for Cube's agent. CI runs in preview |
| Run a query | `sql_query`, one read-only statement on your database | `runQuery`, Cube SQL on semantic views, 100 rows a page |
| Charts and dashboards | `render_chart` and `render_dashboard`, in the AI client | Reports and dashboards saved in Cube, drawn in ChatGPT |
| Who can connect | Unlimited users, your customers' users included | People with a Cube user, Viewer role or higher, or a backend app acting as one |
| Needed first | A data source: your database, or the sample data | A semantic model in Cube Cloud |
| Databases and warehouses | 7, plus PostHog | 29 |
| Tools for the agent | Up to 17 | 30, of which 16 show only to roles that can edit the model |
| How you pay | For each call, no seats | For each developer, plus explorers and viewers on Premium, compute and AI tokens |
| Free plan | 3,000 calls a month | 1,000 requests a day, developer seats only |

## When Cube is stronger

- You want metric definitions enforced: the agent names a measure and Cube writes the warehouse SQL. Bonnard gives the agent the metric's SQL and does not rewrite the query.
- You keep context in git: the model, rules, certified queries and eval questions are files in one repository, and a pull request can review each change ([development mode](https://docs.cube.dev/docs/data-modeling/dev-mode)).
- You want agents to change the model: Developers and Admins edit model files on a development branch through MCP, then commit and merge.
- You want evals that compare result rows with a correct query, and that can block a pull request once Cube turns on the CI preview.
- Your data is in a source Bonnard does not connect to: Cube lists 29, among them Snowflake, Databricks, ClickHouse and Trino.
- Your team works in ChatGPT: since 13 Nov 2026 Cube draws charts and dashboards there and downloads a chart as a PDF ([Cube in ChatGPT](https://cube.dev/blog/cube-in-chatgpt), [changelog](https://cube.dev/changelog)).

## Add each server to your AI client

**Bonnard.** Create a workspace, add Bonnard to your AI client, then add a data source ([quickstart](/docs/get-started/quickstart)). Use the install links for Claude, Cursor and ChatGPT in the console at `app.bonnard.ai`, or add this URL.

```text
https://app.bonnard.ai/mcp
```

**Cube.** The MCP server is on all plans. In Claude, select **Customize > Connectors > Browse connectors** and search for Cube. In ChatGPT, select **Customize > Plugins** and install the Cube plugin described as "Query governed business data". Codex has a Cube plugin too. On a team or enterprise plan of Claude or ChatGPT, an owner or admin may need to turn Cube on first. Other clients use this URL, which serves all accounts and regions.

```text
https://cubecloud.dev/mcp
```

**Admin > MCP Server** in Cube sets the default deployment, limits the deployments that clients can reach and turns off single tools.

Verified 16 Nov 2026 against Cube's [MCP docs](https://docs.cube.dev/docs/integrations/mcp-server), [pricing](https://cube.dev/pricing), [pricing docs](https://docs.cube.dev/admin/account-billing/pricing), [AI tokens](https://docs.cube.dev/admin/account-billing/ai-tokens), [roles](https://docs.cube.dev/admin/users-and-permissions/roles-and-permissions), [AI overview](https://docs.cube.dev/admin/ai), [rules](https://docs.cube.dev/admin/ai/rules), [certified queries](https://docs.cube.dev/admin/ai/certified-queries), [AI context](https://docs.cube.dev/docs/data-modeling/ai-context), [memory](https://docs.cube.dev/admin/ai/memory-isolation), [evals](https://docs.cube.dev/admin/ai/evals), [Chats History](https://docs.cube.dev/admin/monitoring/chats-history), [Query History](https://docs.cube.dev/admin/monitoring/query-history), [Usage Analytics](https://docs.cube.dev/admin/monitoring/usage-analytics), [development mode](https://docs.cube.dev/docs/data-modeling/dev-mode), [data sources](https://docs.cube.dev/admin/connect-to-data/data-sources), [agent skills](https://docs.cube.dev/docs/integrations/agent-skills), [run-query skill](https://github.com/cube-js/cube-agent-skills/blob/main/skills/cube-run-query/SKILL.md), [changelog](https://cube.dev/changelog) and [Cube in ChatGPT](https://cube.dev/blog/cube-in-chatgpt).