Use Tools in Workbench

Use Tools in Workbench

Tools let an AI model do more than write a response.

In Feluda Workbench, a tool can give the model an approved way to retrieve information, work with a file, write to the Journal, use a connected service, or complete another supported action.

You decide which tools are available during the conversation.

This guide explains how to:

• find available tools;
• enable only the tools you need;
• write a clear instruction;
• review tool activity;
• confirm that an action completed; and
• use tools safely.

What a tool is

A tool is a specific capability that an AI model can call while working on your request.

Without a tool, the model can only respond using the information available in the conversation and its own model knowledge.

With a tool, the model may be able to:

• search an approved source;
• read or write a supported file;
• create a Journal entry;
• retrieve live information;
• use a connected service;
• look up technical details; or
• perform another action included with a Gene or connection.

The exact tools available depend on your Feluda setup.

Where Workbench tools come from

Tools can become available through:

• installed Genes;
• supported MCP connections;
• built-in Feluda capabilities; or
• other configured connections.

A Gene may include one or more tools for a particular purpose.

After a Gene is installed and synced, its tools can appear in the Workbench tool selector.

If you do not see the tool you expect, confirm that the related Gene or connection has been added and configured correctly.

Before enabling a tool

Review what the tool does before making it available to the model.

Check:

• the purpose of the tool;
• the type of information it may receive;
• whether it connects to an outside service;
• whether it can change or create information;
• whether it needs an account, key, or private setting; and
• whether its result needs human review.

Start with a non-sensitive example when using a tool for the first time.

Do not enable a tool only because its name sounds useful. Read the Gene, connection, or tool description first.

Step 1: Open Workbench

Open Workbench from the Feluda sidebar.

Select the provider and model you want to use.

Make sure the model is responding normally before adding tools. A short text message is enough to confirm the connection.

For example:

Reply with one sentence confirming that you are ready.

Once the model responds, continue with the tool setup.

Step 2: Open the Tools selector

Select Tools in the Workbench header.

The tool selector shows the tools currently available in your Feluda setup.

The list may include tools from several installed Genes or connections.

Read the tool names and descriptions carefully before selecting anything.

Step 3: Enable only the tools you need

Turn on the tool or tools required for the current task.

For example, if you want the model to create a Journal entry, enable the relevant Journal tool.

If the task only needs one tool, enable one tool.

Keeping the selection small makes it easier to:

• understand what the model can do;
• review which action was taken;
• identify the cause of an error; and
• reduce unnecessary access.

Disable tools that are unrelated to the task.

Step 4: Give a clear instruction

Tell the model what result you want and when the tool should be used.

A useful instruction should explain:

• the task;
• the information to use;
• the action the tool should perform;
• the expected result; and
• any limits the model should follow.

For example:

Use the Journal tool to create an entry titled "Project Review".

Include:
1. a short summary;
2. the current blockers; and
3. the next actions.

Use only the notes below.
Do not add missing details.

Notes:
[Paste the notes here.]

This is clearer than:

Save this.

Ask for an action directly

When you want the model to use a tool, state the action clearly.

For example:

Use the enabled Journal tool to create the entry.
Do not only show me the text in the conversation.

This helps distinguish between:

• preparing content for you to review; and
• asking the tool to perform the actual action.

The model may still decide that a tool is not needed when the request can be answered directly. Review the Activity log to confirm what happened.

Step 5: Watch the conversation

The model may first explain what it is doing, request missing information, or call the enabled tool.

Read any confirmation or question before continuing.

If the model asks for information that the tool requires, provide only what is necessary for the task.

Do not provide passwords, API keys, or access tokens inside the conversation.

Use the intended provider, Secrets, Gene settings, or connection setup area for private values.

Step 6: Review the Activity log

Open the Activity log after the model attempts to use a tool.

The Activity log can help you see:

• which tool was selected;
• when the tool was called;
• whether it completed successfully;
• whether it returned a result; and
• whether an error occurred.

This is the most reliable way to confirm that a tool call was attempted.

Do not rely only on a message from the model saying that the action was completed.

Step 7: Confirm the result

When possible, check the result in its destination.

For example:

• open the Journal to confirm that an entry was created;
• open the expected file to confirm that it was written;
• check the connected service to confirm that the requested change exists;
• compare retrieved information with the original source; or
• review the returned data for missing or incorrect details.

A successful tool call does not guarantee that the result is complete or suitable for use.

Review the outcome yourself.

A simple Journal example

Imagine that you want to save a meeting summary.

First, enable the appropriate Journal tool.

Then enter:

Use the Journal tool to create an entry titled "Weekly Team Meeting".

Write:
* a summary of no more than 100 words;
* the decisions;
* the action items and owners; and
* the unanswered questions.

If an owner or deadline is missing, write "Not provided."

Meeting notes:
[Paste the notes here.]

After the model responds:

1. open the Activity log;
2. confirm that the Journal tool was called;
3. check whether the call succeeded; and
4. open the Journal to review the entry.

A retrieval example

Some tools can retrieve information from an approved source.

After enabling the relevant tool, you could write:

Use the enabled tool to retrieve the latest available information about
[topic].

Return:
1. a short summary;
2. the main facts;
3. the source details returned by the tool; and
4. any uncertainty or missing information.

Review whether the result came from the expected source and whether it is current enough for your task.

Use tools one step at a time

A task may require more than one tool.

Begin with one tool when possible.

For example:

1. retrieve the information;
2. review the retrieved result;
3. ask the model to summarise it; and
4. enable a second tool only when you are ready to save or send the result.

This makes the process easier to understand and reduces the chance of an unintended action.

When you later turn the task into a workflow, separate actions can become separate, reviewable steps.

Understand read and write actions

Tools can have different effects.

A read action retrieves or views information.

A write action creates, changes, sends, or removes something.

Write actions deserve extra care because they may affect a file, account, Journal, or connected service.

Before requesting a write action:

• confirm the destination;
• review the content;
• check the name, title, or recipient;
• understand whether the action can be reversed; and
• confirm that the selected tool is the correct one.

When the result matters, ask the model to prepare a draft first. Review the draft before asking the tool to perform the final action.

Keep the task within the tool's purpose

Use each tool only for the type of work it was designed to perform.

A tool may fail or produce an unexpected result when the request falls outside its supported actions.

If the tool description is unclear:

• review its Gene page;
• review its settings;
• try a simple request;
• check the Activity result; and
• avoid sensitive information until you understand its behaviour.

Use Gene prompts when available

Some Genes include prompts designed for their tools.

Gene prompts may appear in the Workbench prompt selector.

A prompt can help you:

• use the correct instruction structure;
• provide the required information;
• choose a supported action; and
• receive a result in a useful format.

Review and edit the prompt before sending it.

A ready-made prompt is a starting point, not a replacement for checking the task and source information.

Configure required Gene settings

Some Gene tools need settings before they can be used.

These settings may include:

• an account connection;
• a preferred option;
• a source selection;
• an access key; or
• another value required by the connected service.

Open the Gene configuration area and complete the requested settings.

Store private values only in the protected field intended for them. Do not paste them into Workbench.

If a tool appears but repeatedly fails, incomplete settings may be the cause.

Use tools with local models

A local model can use tools in Workbench when the model supports the required tool use and the tool is available.

Keep the local model application running throughout the conversation.

Remember that a local model does not make every tool local.

If the enabled tool connects to a website, online service, or remote data source, the information needed for that action may leave your computer.

Review the tool separately from the model.

Protect sensitive information

Before giving a tool access to sensitive information, check:

• whether the information is necessary;
• where the tool sends or stores it;
• whether an external service is involved;
• whether the result will be saved;
• whether the action can be undone; and
• whether a less sensitive example can be used first.

Remove details that are not needed for the task.

Do not place private credentials inside a conversation, source document, or prompt.

Review tool permissions

Some tools may have limits on what they can access or where they can act.

These limits may apply to:

• websites;
• file locations;
• network addresses;
• ports;
• actions;
• parameters; or
• connected accounts.

A denied action may mean that the request is outside the tool's allowed boundaries.

Do not expand access automatically. First confirm why the action is needed and whether it is appropriate.

If the model does not use the tool

Check that:

• the correct tool is enabled;
• the selected model supports tool use;
• the instruction asks for an action the tool can perform;
• the required information is present;
• the Gene or connection is configured; and
• no required setting is missing.

Make the instruction more direct.

For example:

Use the enabled Journal tool to create the entry.
Do not only draft the entry in the conversation.

Then review the Activity log again.

If the tool is missing

A tool may not appear because:

• its Gene has not been added;
• the Gene has not been synced;
• the required MCP connection is not active;
• Feluda needs to be refreshed;
• the tool is not included in the installed package; or
• the connection setup is incomplete.

Review the Gene or connection guide and confirm that the tool should be available in Workbench.

If the tool returns an error

Read the Activity entry and any message returned by the tool.

Common causes include:

• missing settings;
• an expired or incorrect access key;
• unavailable source information;
• an unsupported action;
• a connection problem;
• a denied permission; or
• incomplete input.

Correct one issue at a time, then test the tool with a small request.

Avoid repeating a write action until you have confirmed that the first attempt did not already complete.

If the result is incorrect

A tool can complete successfully but still return information that does not answer your question.

Check:

• whether the instruction was specific enough;
• whether the correct tool was selected;
• whether the source was appropriate;
• whether the requested filters were clear;
• whether the result is current; and
• whether the model interpreted the tool output correctly.

Ask the model to separate the raw tool result from its own summary when you need a clearer review.

Disable tools when the task is complete

When you finish the task, turn off tools that are no longer needed.

This keeps the next conversation focused and reduces unnecessary access.

Start a new conversation when you move to unrelated information or a different set of tools.

Move a successful tool task into Studio

A tool-assisted Workbench task may be suitable for a workflow when:

• the goal is clear;
• the required tool is known;
• the input follows a regular pattern;
• the expected result is easy to review; and
• the task needs to be repeated.

Test the task several times in Workbench first.

Then create a workflow in Studio that gives the relevant AI step access to the required tool and defines what should happen with the result.

Keep write actions and important decisions reviewable.

A safe first tool session

For your first tool-assisted conversation:

1. choose a non-sensitive task;
2. enable one tool;
3. read the tool description;
4. give a direct instruction;
5. review the Activity log;
6. confirm the result in its destination; and
7. disable the tool when finished.

This gives you a clear understanding of how Workbench, the model, and the tool work together.