How to use Claude with Fishbowl
Short answer
Claude reaches Fishbowl through an MCP server that wraps the Fishbowl REST API or a read-only database user. You can run Claude in the Anthropic API, through AWS Bedrock or Google Vertex, or swap in a local open-weight model if the data cannot leave your network.
Claude is a good fit for the messy parts of Fishbowl work: reading a vendor's long email chain against a PO, reasoning across a multi-level BOM, or explaining why a number on a report does not match the screen. It is less about replacing Fishbowl and more about a capable reader sitting beside it.
The connection question is the same as for any on-prem system. Claude cannot see inside your network. The standard bridge is a Model Context Protocol server, a small program that exposes defined tools such as get_part, list_open_pos or get_sales_order. Claude desktop and other MCP-aware clients can call it.
The second question is where Claude itself runs. That is a policy choice, and the answer differs between a distributor with ordinary stock data and a shop making controlled parts.
What teams use it for
Reading vendor emails against POs
Paste or forward a vendor confirmation and ask Claude to compare quantities, dates and prices with the PO lines pulled through a tool.
BOM explanation and sanity checks
Ask Claude to walk a multi-level BOM, flag components with no vendor or zero cost, and describe what would be affected by a substitution.
Month-end inventory variance notes
Give it adjustment history and have it draft explanations that an accountant can verify before closing.
Long document summarization
Summarize vendor contracts, quality docs or customer specs attached to a part record, with a person checking key terms.
Report logic review
Paste the SQL behind a custom report and ask Claude to explain it and spot joins that double-count.
Help desk for new staff
A tool-enabled Claude answers how-do-I questions about receiving, picking and manufacture orders using your own written procedures.
Fishbowl integration surface
- Fishbowl REST API with an approved integrated app and token auth: the cleanest route for an MCP tool that reads parts, orders and inventory. Endpoint coverage varies by version.
- Read-only MySQL access: the fastest way to answer analytic questions, especially across tables the API does not cover.
- Legacy JSON API: works for older servers if you must support them behind the same MCP tool signatures.
- Fishbowl report exports and CSV: handy as static context files for questions that do not need live data.
- Part-level attachments and notes: useful context for summarization, accessible through the database or API depending on version.
How it works
- 1
Write the MCP server
Use an MCP SDK in Python or TypeScript. Define four to eight read-only tools with typed inputs and small JSON outputs.
- 2
Authenticate to Fishbowl inside the server
Store the integrated app credentials or database password in the server's environment. Claude never sees them.
- 3
Choose a transport
stdio works when the client and server share a machine. For team use, run the server over HTTP inside your network with its own authentication.
- 4
Choose where Claude runs
Options include the Anthropic API, Claude via AWS Bedrock, or Google Vertex AI. Each has its own data handling terms; confirm them with the provider for your region and plan.
- 5
Write operating instructions
A short system prompt: units of measure, location naming, what costs may be shown to whom, and to cite the part number used in every answer.
- 6
Test with real questions
Build a list of fifty questions your staff actually ask and check answers against Fishbowl. Fix tool gaps rather than prompting around them.
- 7
Decide on writes later
Add draft-creating tools only after reads are solid, with approval gates.
Example flows
PO confirmation check
Trigger: Buyer receives a vendor acknowledgment email.
- Buyer pastes the email into Claude.
- Claude calls get_purchase_order with the PO number.
- It compares lines, quantities, unit costs and dates.
- It lists discrepancies and drafts a reply.
Result: Discrepancies are caught before goods arrive.
Why is this part short
Trigger: A planner asks about a part shortage.
- Claude calls get_part_availability.
- It calls list_open_po_items and list_open_manufacture_orders for the part.
- It explains on-hand, committed, on-order and the earliest relief date.
Result: A one-paragraph answer with the numbers behind it.
Keeping data on-prem
If data cannot leave your network, you have two choices. First, keep Fishbowl and the MCP server on-prem and use a cloud route such as Bedrock or Vertex inside a controlled cloud account, sending only the small tool outputs rather than bulk data. Second, run an open-weight model locally with vLLM or Ollama and the same MCP tools, which many MCP clients can use with a compatible model. The tool design is identical either way, so you can start in the cloud and move local later without rewriting the integration.
Security checklist
- Run the MCP server with a read-only database user or a scoped API account.
- Return only the fields needed for each tool, and remove cost fields for roles that should not see them.
- Require authentication on any HTTP-transport MCP server and keep it off the public internet.
- Log tool calls with user, tool, arguments and time.
- Review provider data retention and training terms before sending production data.
Ways to get it done
Claude with exports only
Effort: Minutes.
Ad hoc analysis and writing, no integration.
Self-built MCP server over the REST API
Effort: Days to a couple of weeks.
A team with a developer and a handful of live questions to answer.
Partner build with Netray, cloud or local model
Effort: Scoped pilot, then rollout.
Regulated data, role-aware access, and proper audit and evaluation.
Pitfalls
- !Returning huge result sets to the model, which wastes context and degrades answers. Paginate and summarize.
- !Letting Claude guess at part numbers instead of forcing a lookup tool call.
- !Skipping unit of measure and location conventions in the instructions.
- !Treating a narrative explanation of variances as fact without a person verifying the underlying numbers.
- !Running the MCP server with a database admin account.
Frequently asked questions
Does Claude have a Fishbowl connector?
I am not aware of an official one, so check current connector directories. The practical route is a custom MCP server you or a partner build over Fishbowl's REST API or a read-only database user, which Claude and other MCP clients can use.
Can Claude run entirely inside my network?
Claude models are provided through Anthropic and cloud partners, so a fully air-gapped setup uses a local open-weight model instead. Keep your MCP tools identical and the model becomes swappable, so you can pick based on policy.
Is Claude better than ChatGPT for Fishbowl?
Both can work. The integration pattern is the same, and MCP lets you use either. Choose on data terms, long-document handling and cost for your use case, and test with your own questions rather than relying on general claims.
Can Claude change data in Fishbowl?
Only if you build write tools. Start read-only. If you add writes, have them create drafts that a person approves in Fishbowl, and use a separate scoped account so every change is attributable.
What do I need to build the connection?
A Fishbowl integrated app approved by an admin or a read-only MySQL user, a machine on your network to host the MCP server, and a developer comfortable with an MCP SDK. Most of the effort is in tool design and testing, not the plumbing.
Want Claude working on your Fishbowl?
A free 30 minute assessment: we map the integration path for your version and setup, and where it should run.
Get a free assessment