This guide describes the 1.0.0 workbench shown in the product tour. Remote tool steps use a
saved local connection ID; credentials stay outside the .aether file.
Set up a connection
- Open MCP Connections from the activity bar or command palette. Add an HTTP server address, choose a compatible remote entry from the public Registry, or configure a local stdio executable. Registry search does not install or authorize an app.
- Give the connection a name and stable ID. For HTTP, choose no credential, bearer token, custom header or OAuth account sign-in. For stdio, provide the executable, literal argument array, optional working directory and explicit environment values. Bandura uses no shell.
- Save. For OAuth, Sign in with browser opens the server’s authorization page; return to Bandura after granting access. Reconnect, disconnect or cancel sign-in from the same page. Token scope/expiry is shown when reported; protected refresh tokens stay on this device.
- Check tools lists the exact names and schemas without invoking a tool. Inspection starts the configured executable for stdio and stops it afterward. Stop inspection cancels it; another check starts a fresh process. Check all checks at most three saved apps concurrently.
Search connections by name, ID or address, then search tools by name, description or input field. A successful catalogue is cached with its timestamp. Refresh keeps your authored inputs, reports schema changes and leaves cached discovery visible on failure. Check again before execution: listing a tool does not establish permission to run it or the correctness of its data.
Saved account, credential or address changes invalidate the relevant catalogue. Credentials do not carry to another endpoint. Copy setup manifest shares the ID and authentication shape without saved tokens, URL query values or local executable/environment details. A teammate must set up their own connection; review names and paths before sharing.
Configure a tool step
Open a workflow and choose Add to open workflow beside an inspected tool, or add an
mcp-tool node and open Configure. Search the saved app and exact tool. Type to filter,
use arrows and Enter to select; Escape keeps the previous selection. No tool runs during search.
A missing connection offers explicit local mapping or Manage connections; Bandura does not
guess which account should replace it.
The form supports nested objects and arrays, numbers, booleans, enums, required/optional values
and null. Use default explicitly applies a server-declared default. Variable accepts a
whole ${{ }} expression with its runtime type; JSON handles unions or unsupported form
constructs. Incomplete drafts survive node/tab changes. Values the form does not represent remain
in Source for review. Inputs are checked again against the current schema after variables resolve;
unsupported local dialects say Checked by server only.
- id: find_customer
type: mcp-tool
config:
connection: customer-app
tool: search_customers
arguments:
email: "${{ variables.email }}"
output: customer
next: check_customer
The tool name/inputs above are illustrative: use what your server actually exposes.
Copy output reference gives ${{ variables.customer }} for later steps. Result shows the
raw MCP content and the assigned output separately. Structured false, zero, arrays and null are
preserved; content blocks are the fallback. Tool errors fail the step but retain returned evidence.
Use the connection in chat
Open Chat → Context and enable the specific connection. Chat can inspect exact schemas, list/read negotiated resources and get prompts. Returned data may be sent to your selected AI provider; a connection checkbox does not authorize tool execution. Remove it to narrow future turns. Flow, selected step/result and automatic workspace retrieval have separate controls.
Chat can validate and propose a workflow using the inspected tool. Review the file diff before Apply. To run a tool, the desktop’s shared action policy applies to direct chat calls and workflow/subflow runs: Review each action by default, or explicitly saved allow/deny with an optional exact tool allowlist. Server read-only hints are advisory, not permission. The review binds the actual connection, current schema and resolved arguments. If those change, the call fails for fresh inspection/review. Stop cannot undo a request already sent; check a write’s outcome before retrying a timeout.
Run, limits and sharing
Inspect the run result/history and add assertions before another external action. MCP replies
and catalogues are bounded to 1 MiB, lists to 500 catalogue entries, inspection to ten seconds and calls to
30 seconds by default (timeoutMs). Negotiated input requests show a separate form/browser
step. Supported legacy task servers show progress with bounded polling; Stop requests cancellation.
Server cancellation is not a guarantee of rollback.
Saved connection IDs, OAuth accounts, stdio setup and review dialogs are desktop features.
Headless CLI/MCP-host flows use serverUrl and environment-backed headers. Share workflow files,
connection setup instructions and required variables, without protected credentials. The separately
versioned published CLI may not include the latest development node features.