connect

Files (1)

• SKILL.md

SKILL.md (first lines)

# Summation Connect

Help the user get a **live data source** into Summation, then make its tables analyzable under **human-readable names**.

## Product knowledge (required — live, not baked in)

1. **Fetch `https://docs.summation.com/llms.txt`** before answering which sources are supported.  
2. Open the matching connector page from that index (e.g. Postgres, Snowflake, BigQuery).  
3. Hosted multi-engine services (RDS, Cloud SQL, Azure Database, …) are **not** a connector by themselves — ask which engine (Postgres, MySQL, SQL Server, …) before mapping to the index. Neon and similar Postgres-only hosts can map to Postgres once confirmed.  
4. See `../api/references/product.md` for lookup rules only — **not** a static feature list.

Never invent “unsupported” after a failed URL, truncated API dig, or homepage marketing. If docs can’t be reached, say so and retry — don’t guess.

## Customer path for new connections (default)

**Passwords and secrets never go in chat or MCP tool arguments.**

1. Confirm the source type (supported list).  
2. Collect **non-secret** fields only; echo for confirmation.  
3. Guide the user:

   > Open Summation → **Connections** → **Add connection** → pick **\<Source\>**.  
   > Enter the settings we listed. Put the **password only in the browser**.  
   > Click **Test**, then save. Tell me when it’s done.

4. Wait. Do **not** dig through OpenAPI or guess config key names in chat.  
5. When they say done: MCP `list_data_connections` → `test_data_connection` → browse → **attach with names** (below).

If they already have a connection: skip create; go to attach.

(If MCP later exposes a safe create path that never puts secrets in the model transcript, you may use it — still never paste secrets into chat.)

## After the connection exists (MCP)

1. `list_data_connections` — find by name.  
2. `test_data_connection` — pass/fail in plain language.  
3. `browse_connection_resources` — preview of what can be attached.  
4. **`attach_connection_datasets` — always set a clear `name`** (source table name: `customers`, `invoices`, …). Never leave auto names like `pg_…_dataset_2`.  
5. If rename/detach tools exist, use them to fix mistakes; otherwise guide the user in Connections.  
6. Verify with `list_connection_datasets` (friendly names). Attach to the project catalog when needed.  
7. Only then: “Connected **X**. Tables ready: **a, b, c**.”

## File / CSV import

Shell available: prefer **`sumcli tables import`** (the direct pipeline; needs the CLI session — `../api/references/sumcli.md`). Shell-less: `import_file_to_table` (and related file tools) via MCP. Confirm with catalog/list after import.

If the MCP import fails with 409 "not enabled for this workspace" (`feature_not_enabled`): that gates only the agent-workflow route. The **direct pipeline** still works on both surfaces — `sumcli tables import`, or over MCP: `create_asset_upload_url` → PUT the bytes to the presigned URL → `preview_import_asset` → `create_table_import`. Use whichever surface you are already on; no admin action needed. Other failures: say so plainly and offer to connect a warehouse, or retry after fixing the issue — no internal tool names.

## Hard bans (user chat)

- OpenAPI archaeology or multi-variant key guessing  
- Secrets in MCP args or chat (salvage: web app + rotate if already pasted)  
- Secretless orphan connections  
- Claiming success without a final list/test  
- Dumping paths, schema JSON, or internal tool ids  

## Auth

MCP auth failure → the `signin` skill.

From summationai/summation-plugin, MIT.

More from summationai/summation-plugin