Home

Tool Calling in Chat

Bodhi App's chat UI is agentic: when you connect an MCP server, the chat input grows a tool selector, the model can decide to invoke tools, and the resulting tool calls and results are rendered inline as the conversation unfolds.

This page covers the chat-side experience. For setting up an MCP server (auth, base URL, allowed tools), see MCPs → Setup.

What you need

Before tools become useful in chat:

  1. The model you select must support tool/function calling. Most modern instruction-tuned models do; reasoning models like the GPT-OSS series do too.
  2. At least one MCP server must be configured and enabled. See MCPs → Overview.
  3. The MCP server's tool list must be discovered — Bodhi caches this; older instances may need a refresh from the MCPs page.

If any of these is missing, the tool selector still appears but will be empty or dimmed.

The MCPs popover

Look for the plug icon in the chat input row. Click it to open the MCPs popover.

The popover lists every MCP instance configured for your account. For each:

  • A checkbox enables or disables all tools from that server.
  • An expand chevron reveals the per-tool toggles.
  • A counter shows enabled tools out of total (e.g. 3/5).
  • A small spinner appears while the chat UI is connecting to the server.

When at least one tool is enabled, a numeric badge appears on the plug icon showing how many tools are active across all servers.

Why a server might be unavailable

Servers that you cannot use right now are dimmed, with a tooltip explaining why:

  • Disabled by administrator — the platform admin disabled this MCP server entry. Ask them to re-enable it.
  • Disabled by user — you've turned this MCP off on the MCPs page. Re-enable it there.
  • Tools not yet discovered — Bodhi hasn't been able to enumerate this server's tool list. Open the MCP Playground and run a refresh.

Per-conversation tool selection

Tool selections are scoped per chat:

  • The set of enabled tools you pick is saved alongside the conversation.
  • Reopening an old chat restores its exact tool selection — past conversations replay with the same toolset they were generated under.
  • New chats start with the last-used selection, so you don't have to retick the same tools every time.
  • Switching models mid-chat does not clear your selection, but if the new model doesn't support tools, the popover becomes inert.

The selection lives in the same browser-local IndexedDB store as your chat history. There is no server-side per-conversation tool config.

The agentic loop

Once tools are enabled and you send a message, here's what happens:

  1. You send your message.
  2. Bodhi sends the message plus the JSON schema of every enabled tool to the model.
  3. The model either replies normally, or returns one or more tool_call requests.
  4. For each tool call, Bodhi:
    • Locates the right MCP instance from the tool name (Bodhi prefixes tool names with the MCP slug to keep them unique across servers).
    • Calls the tool via the MCP client, passing the model's arguments.
    • Captures the tool's response (or error).
  5. Bodhi sends the tool results back to the model.
  6. The model continues — it may answer the user, or call more tools.

Steps 3–6 can repeat. The maximum number of rounds is bounded by the Max Tool Iterations setting (default 5; configurable in the Settings panel). Hitting the limit forces the model to produce a final answer instead of calling another tool.

When the model requests multiple tool calls in one turn, Bodhi executes them in parallel for lower latency.

How tool calls render in the conversation

Each tool invocation appears as a collapsible card embedded in the assistant message:

  • Header shows a wrench icon, the tool name, and the source MCP slug.
  • Status badge updates live:
    • Calling... (blue, spinner) while the tool is running.
    • Completed (green, check) on success.
    • Error (red, cross) on failure.
  • Arguments (collapsible): the JSON arguments the model produced, pretty-printed.
  • Result (collapsible): the tool's response, also pretty-printed. Errors are tinted red.

Cards auto-expand while the tool is executing so you can watch arguments and results land. They collapse once execution finishes — click the header to re-expand any card. If the result is JSON it's formatted; otherwise it's shown verbatim.

Tool calls and their results are saved into the chat history so re-opening a conversation shows the full audit trail.

Error states

The most common failures and what they mean:

  • Tool call returns Error: the MCP server itself reported a failure. The error text is shown in the result panel and forwarded to the model on the next turn — the model can choose to retry, call a different tool, or apologize.
  • MCP unreachable: if Bodhi can't connect to the server at all, the call fails with a connection error before the tool runs. Check the server's URL and auth on the MCPs page.
  • Auth expired: for OAuth-backed MCP servers, an expired token surfaces as an error in the result. Re-authorize the server from the MCPs page and resend the message.
  • Rate-limited or timed out: the model may retry the tool call on its next turn. If it loops, lower the Max Tool Iterations value or disable the offending tool.
  • Model ignored the tools: not every model is good at choosing tools. Try a stronger model, or rephrase your message to hint at which tool to use.

Thinking models with tools

Reasoning-capable models that also support tools (e.g. GPT-OSS-120B, certain Claude variants) interleave thinking blocks with tool calls. The chat UI renders both inline:

  • A collapsible thinking block above or between tool cards shows the model's reasoning trace.
  • Tool cards appear in execution order.
  • The final assistant message follows once the model decides it has enough information.

Tips

  • Start small. Enable just the tools the model needs for the task. Sending fifty tool schemas to a small model wastes context and confuses tool selection.
  • Use stop sequences sparingly. A stop sequence that lands inside a tool-call payload will truncate the call. See Parameters and System Prompt.
  • Test tools first. Use the MCP Playground to call a tool by hand before relying on it in chat. If it doesn't work standalone, the model can't fix it.
  • Use the system prompt to guide tool use. A line like "Prefer the search_docs tool before answering from memory" materially changes selection behaviour.