# Chat

The AI assistant for your entire workspace.

Chat is SteelEngine's AI assistant. There is one Chat, scoped by where you open it. A Chat started from Home or a contextual tray works across your whole workspace. A Chat opened inside the workflow editor, from the **Assistant** tab of the editor's right panel, is focused on that workflow (see [Chat in the Workflow Editor](/chat/workflow-editor)). The editor's floating **Chat** panel is not the assistant. It is the test panel that runs the workflow you have open.

A workspace-wide Chat can build workflows, edit existing resources, analyze tables, create files, search knowledge bases, schedule recurring tasks, and open the relevant resource on the right side of the conversation.

**Naming standard:** Chat is the canonical name for SteelEngine's adaptive assistant in every assistant scope. Use "workspace-wide Chat" or "Chat in the workflow editor" when the scope matters. Mothership, Copilot, and title-cased Workspace Assistant are retired product labels that remain only in compatibility identifiers and legacy URLs. A workflow's generated chat interface is **Hosted Chat**, a separate deployment surface.

Use Chat for adaptive work that benefits from reasoning against current workspace context. Use a deployed Workflow for a defined process that must run repeatedly and predictably. Chat can build, manage, and debug that Workflow.

## What You Can Do

[Workflows](/chat/workflows)

Build, edit, run, debug, deploy, and organize workflows.

[Research](/chat/research)

Search the web and read pages through search integrations, and write research reports.

[Docs](/chat/docs)

Upload, create, and edit workspace docs, and chart data inline.

[Tables](/chat/tables)

Create tables, query rows, and update records.

[Automation & Configuration](/chat/tasks)

Schedule tasks, take direct actions, connect integrations, and configure workspace settings.

[Knowledge Bases](/chat/knowledge)

Create knowledge bases, add documents, query content, and connect them to workflows.

## How It Works

Chat uses the shared platform tool gateway directly. [SteelEngine MCP](/mcp/steelengine) exposes that gateway to external clients; Chat does not make an HTTP request to its MCP endpoint. The catalog has one `steelengine_*` tool per resource area: workflows, tables, docs, knowledge bases, scheduled tasks, credentials, environment variables, MCP servers, skills, memory, and logs. Separate tools search the integration catalog and execute the matching integration with its advertised input schema. Chat does not receive the contents of your workspace up front. With each message it receives the current date and a bounded inventory: the names and IDs of the connected accounts, tables, knowledge bases, workflows, and other resources you can read. Your permissions are checked again for every message, so a resource you lose access to is left out of the next one. Chat looks up anything else, so you can refer to resources by name:

* "Run the invoice workflow"
* "Analyze the leads table and open it"
* "Add this PDF to the support knowledge base"
* "Schedule the morning CRM sync for 8am every weekday"

When you open Chat from a page that shows data, such as a table, a log, or a workflow, Chat receives the on-screen view as context (up to 12,000 characters). Every tool call appears in the thread as a collapsible row that names the action. Expand the row to see its parameters and result. For background assistant work, create a scheduled [Sub Agent](/sub-agent). Automatic delegation to separate Sub Agents during an interactive response is currently disabled.

## Approvals for Destructive Actions

Some operations pause the response and ask for your approval. These are the operations the tool catalog marks as destructive or confirmation-required, such as deleting a resource, deploying or redeploying a workflow, or reverting a deployment version. The thread shows an in-chat prompt ("… needs your approval to run.") with **Approve** / **Deny** buttons. Denying prevents the call and ends tool execution for that response; Chat does not retry it with different arguments. Workspace invitations identify the invitee and role in the approval prompt.

Integration tools that write (send, create, update, delete, and similar verbs) ask the same way. Read-only integration tools (get, list, search, and similar) run without asking. A test run of a saved workflow draft asks too when any block in the draft could change something; see [Creating Workflows](/chat/workflows#creating-workflows). Resuming a run that is paused at a Human in the Loop step also asks, and the prompt shows the answer Chat will give; see [Resuming a Paused Workflow](/chat/workflows#resuming-a-paused-workflow). External MCP clients use `confirm: true` for confirmation-required platform operations. Scheduled Sub Agents run without interactive approval prompts; see [their execution rules](/sub-agent#authorization-and-actions).

## Connecting Integrations

When Chat needs a Slack or Salesforce connection, its connection prompt lets you choose the hosted provider or a permitted [custom app](/credentials/custom-apps). The selected account stays bound to that app. Personal Slack connections use your own user token; workspace bot installations follow their credential permissions.

## Adding Context

There are three ways to bring workspace objects into the conversation:

* Open the picker with the `+` menu or by typing `@` (both open the same picker)
* Drag resources from the sidebar
* Open a resource in the right-side panel

Chat can also open resources automatically when it creates or modifies them.

The active resource tab takes priority over other open tabs. References carry resource IDs, including workflow branch and selected-block IDs where relevant. An attached past chat supplies a bounded transcript from that exact conversation, so chats with the same title stay distinct. Resource details are refreshed under your current permissions; saved on-screen snapshots can contain historical or unsaved content. Unavailable or oversized context is marked rather than silently treated as a complete current copy.

| Resource           | How Chat uses it                                                                                                    |
| ------------------ | ------------------------------------------------------------------------------------------------------------------- |
| **Workflow**       | Reads and edits the canvas, then opens it in the resource panel                                                     |
| **Table**          | Queries, updates, or opens the table editor                                                                         |
| **Docs**           | Reads, edits, previews, or generates workspace docs                                                                 |
| **Knowledge Base** | Searches documents or adds new documents                                                                            |
| **Past chat**      | Reuses an earlier conversation as context                                                                           |
| **Subagent**       | Reads the selected background task's configuration and run history; open it in Automations for its full detail page |
| **Log**            | Reads execution summaries and failure details                                                                       |

## Model and Reasoning Effort

The input has a model picker and, for models that support it, a **Reasoning effort** picker. Only reasoning levels supported by the selected model are offered. Changing models removes an incompatible effort choice, and requests with unsupported combinations are rejected. Scheduled Sub Agents use the same model roster and reasoning rules.

The roster is a curated list of provider models, subject to workspace and organization restrictions, and only shows models whose provider has a key: the workspace's own (BYOK) or the deployment's. Responses use the applicable BYOK key or the deployment's provider key (see [Bring Your Own Key](/credentials/bring-your-own-key)).

The roster includes Claude Fable 5.2, Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6.2 Astra, GPT-6.2 Sol, GPT-6 Astra, GPT-6 Sol, GPT-6.1 Terra, GPT-6 Luna, GPT-6.1 Sol, Gemini 4 Argon, Gemini 4 Carbon, Gemini 4 Hydrogen, Grok 4.7, Qwen3.8 27B (Cerebras), and DeepSeek Flash alongside the existing models. Claude Opus 5.5, Sonnet 5.5, Fable 5.1, and Fable 5.2 always use adaptive thinking; **Fast** uses low effort, and **Balanced** uses the model's default effort (medium for Opus 5.5, high for Sonnet 5.5, Fable 5.1, and Fable 5.2).

A saved selection of a model that is no longer offered moves to its closest current model: DeepSeek V4 Flash, `deepseek-chat`, and `deepseek-reasoner` move to DeepSeek Flash, Qwen3 235B moves to Qwen3.8 27B, and retired Claude and Gemini models move to their current equivalents.

If a saved model loses its provider key or is blocked by the deployment, Chat falls back to the workspace's default Chat model. The default is `claude-opus-5` when available, otherwise the first available model in the roster. This also applies in the workflow editor and to scheduled Sub Agents. The selected model must still pass reasoning-effort, provider, and organization access checks; an empty roster needs provider configuration before Chat can run.

Ask "What model are you running?" to get the model name and provider for the current response. Chat receives the model selected by the server on every turn, including after a fallback. Switching models takes effect on the next response; earlier replies can come from a different model. SteelEngine is the product hosting the conversation.

## Memory

Chat keeps long-term memory scoped to you within the workspace. It can save a note under a key and search saved notes later, so preferences and facts carry across conversations. MCP clients reach the same store through `steelengine_memory` (list, search, add, correct, delete).

## Voice Input

When the deployment has speech-to-text configured, the input shows a **Voice input** microphone button. Click it to dictate. Chat appends the transcript to whatever you have typed. Click the button again (**Stop listening**) to end the session. A session lasts at most three minutes. When billing is enabled, voice sessions count toward your usage.

## Layout

A workspace-wide Chat has two panes:

* The chat thread on the left shows your messages, assistant responses, thinking blocks, tool progress, and approval prompts.
* The resource panel on the right opens workflows, tables, files, knowledge bases, past chats, and other resources as tabs.

The panel is resizable. Tabs can be opened, activated, closed, and reordered as the conversation evolves.

## Long-Running Work

A single response is bounded. Chat makes at most 20 model calls and 40 tool calls per response. When it reaches either limit, or four minutes after your message arrived, it uses its last step to report what it finished and what is left, so you can ask it to continue. A tool call still running at that point is stopped, and Chat reports its outcome as unknown instead of finished. A workflow run that Chat had requested but that had not yet been accepted for execution is not started. A run that was already accepted keeps going; open its execution log to see the result. On hosted SteelEngine a response is also capped at five minutes.

While Chat is responding, the send button becomes **Stop**. There is no message queue, so send follow-ups once the response finishes. Answering an approval prompt continues the same response. If that continuation fails, **Retry** sends your answer again and shows any saved result. For an approved draft workflow run, Retry uses the original execution ID. For other approved actions without a saved result, Chat reports that the outcome is unknown. Check whether the action took effect before requesting it again. For recurring work, ask Chat to create a scheduled task.

## Using Chat from the API

Your own code can send messages to Chat with an API key. Call [Send Chat Message](/api-reference/chat/sendChatMessage) (`POST /api/v1/chat`) with a `message` and, optionally, a `workflowId` to use as context. The request waits for the finished reply and returns Chat's text, the tools it called, and a `chatId`.

Send that `chatId` with the next message to continue the same conversation. Chat reuses its workflow and message history, so you do not need to repeat `workflowId`. If you specify a workflow, it must match the conversation's workflow or the request returns `400`. Only chats attached to a workflow can be continued through this endpoint. An unknown or inaccessible `chatId` returns `404`; omit `chatId` to create a new conversation. The older `/api/v1/copilot/chat` path still works and behaves the same.

An API request has nobody to answer an approval prompt, so `workflowRuns` decides which workflows Chat may run. With the default, `safe`, Chat runs a deployment, and runs a saved draft only when no block in it could change anything; it reports any other draft as untested. It does not resume a paused run, because the step the run is waiting at needs a person's decision. Set `none` to keep Chat from running workflows at all, for example when your own code runs what it built. Set `all` to let Chat also run a draft that makes changes and resume a paused run.

Chat acts with your full access, so it needs an unrestricted personal API key with write access to the workflow, in a workspace that allows personal API keys. Scoped, workspace and organization keys are rejected with `403` for now. Requests share the v1 API request limit with the workflows, logs, and audit logs endpoints and return `429` with `Retry-After` when it is exceeded.

## Common Questions

**How is a workspace Chat different from Chat in the workflow editor?**

It is the same assistant, scoped differently. A Chat opened inside the workflow editor is focused on the workflow you have open. A Chat opened from Home or a tray is workspace-wide and can coordinate workflows, tables, files, knowledge bases, scheduled tasks, credentials, integrations, and research from one conversation.

**Can Chat edit an existing workflow?**

Yes. Reference the workflow by name, open it in the resource panel, or add it with @ or the + menu. Chat reads the existing workflow before applying edits.

**Can Chat open tables or files on the right side?**

Yes. When it creates, reads, or updates a workspace resource, it can open that resource as a tab in the right-side panel.

**How long can a task run?**

One response makes at most 20 model calls and 40 tool calls, and on hosted SteelEngine it is capped at five minutes. At a limit Chat reports what it finished and what is left. If the work is larger, continue in a follow-up message. For recurring work, ask Chat to create a scheduled task.

**Can Chat work on multiple messages at once?**

No. Chat processes one response at a time, and the send button becomes Stop while it works. Send the next message after the response finishes.

**Will Chat delete or deploy something without asking?**

No. Destructive and confirmation-required operations show an Approve / Deny prompt in the thread. A denied call never runs.
