Ask Morse

Ask Morse is a chat panel available from every dashboard page — open it to ask about your own agents in plain English instead of building a filter or reading a chart yourself. Morse speaks as your conversational interface in the first person. This chat is read-only: it can inspect supported records and guide you through the product, but cannot open a PR, change settings, apply a fix, or send a message on your behalf.

Ask Morse is in beta. Answers are built from your Morse data and cite it where they can, and an answer whose numbers can’t be tied back to that data is withheld rather than shown. Check important figures against the dashboard before acting on them. Ask Morse is not yet switched on for every workspace; where it is off, the panel can’t answer questions.

What you see

Ask Morse is a panel, not a route — it rides along on top of whatever page you’re already on.

Docked, floating, and mobile

Three layout modes: docked (a 420px right rail, non-modal — the page reflows around it), floating (a centered modal roughly 80% of the viewport, opened with a maximize button), and mobile (a near-fullscreen modal; there’s no useful docked state on a small screen, so the maximize toggle is hidden).

A header carries the “Ask Morse” title plus maximize/minimize and close controls.

Open it from the Ask Morse trigger in the dashboard header, or a keyboard shortcut.

Doc mode and data mode

Two modes decide what Morse can answer, and the switch is automatic — there’s nothing to toggle. On every request, Morse checks whether your org has any trace data yet:

No trace data: Morse answers from a product/SDK docs corpus only — setup and how-it-works questions, not your own data.
Any trace data: Morse gets the full tool set — your traces, agent stats, costs, logs, alerts, and the same investigations and risk analysis the dashboard itself shows.

The mode check runs per request, not once per session — a workspace can flip from doc mode to data mode mid-conversation the moment its first trace lands, with no action from you.

Before you have any data, try a starter prompt from the empty state. Once you have traces, ask things like: “Hey Morse, why did cost spike on the billing agent in the last 24h?”

Starter questions follow the page

The starter questions in the empty state depend on the page you opened Ask Morse from — up to four, limited to what your Beta scope can answer. Before the page is known, the only chip is What can you do?

Page	Starter questions
Home	”What needs my attention first, and who owns it?” · “How are my agents doing this week?” · “Why did my spend change versus last week?” · “Which agents are getting better?”
Cost	”Where is my spend going, by agent and model?” · “Why did my spend change versus the previous period?” · “Which spend should I look at first?” · “Which agents are spending more without getting better?”
Quality	”Is quality dropping on any agent?” · “How many runs have had a sampled answer check?” · “Which agents are getting better?” · “What needs my attention on quality?”
An agent’s page	”How is {agent} doing this week?” · “Why did {agent}‘s latest failed run fail?” · “Where does {agent}‘s spend go?” · “Is quality dropping on {agent}?”

Every other page uses a general set.

Ask about this

On Home, Cost, Quality and an agent’s page, the header’s Ask Morse button opens Ask Morse with that page’s starter questions, and the first one is Ask about this page (on an agent’s page, Ask about this agent). Selecting it puts a question about the page in the composer — for example on Home: “Walk me through Home: what needs my attention first, who owns it, and what evidence is missing.” Nothing is sent until you press send, so you can edit the question first. On other pages, Ask Morse opens with its general starters.

Streaming responses

Replies stream in as Morse works — you see each reasoning and tool step appear in order, then the answer typing in, rather than a spinner followed by a wall of text.

Explain a trace

On a trace’s Run summary, an Explain this run button sends that trace straight to Morse’s explain_agent_dag tool, which classifies its orchestration into one of six known patterns — ReAct, plan-execute-reflect, hierarchical multi-agent, sequential pipeline, map-reduce, or single-shot — or flags it as unrecognized rather than forcing a guess. The explanation opens in a sheet titled This run, explained.

How Ask Morse answers

Before it answers, Morse sorts your question into one of a few paths, so each answer comes from the right source.

Questions about Morse itself (“how do I set up the Python SDK?”) are answered from the Morse docs for the release you’re on. The docs Morse reads are rebuilt when a new release ships, and a saved answer is only reused for the same docs release. These questions are free: they don’t use up your plan’s monthly Ask Morse messages.
Questions about your data (“what did the billing agent cost last week?”) use the same numbers as the dashboard: spend, outcomes, eval scores and attention items are read through the query layer the dashboard uses, not recalculated separately. Each answer ends with a Sources line linking to the dashboard page or trace the figures came from. A question that needs both the docs and your data is answered from both.
Long-running questions (more than 90 days of data, or a wide scan) finish in the background. Morse tells you it’s working and roughly how long it will take, and you can keep using the panel. When the answer is ready you get an in-app notification and then an email, each with a link to the answer and no figures in it. If the question can’t be completed, nothing is delivered and the message is given back. Each organization can queue a limited number of these per hour.
Saved figures show “as of”. To answer common questions quickly, Morse can reuse figures it computed a few minutes earlier. When those figures are more than a minute old, the answer shows when they were computed, for example “As of 4 min ago”. No line means the figures are current.

Questions about your data and long-running questions each use one message from your plan’s monthly allowance. If a number in an answer can’t be tied back to your data, the answer is withheld, as described in the beta note above.

How to use it

Open the panel from the header trigger or the keyboard shortcut; toggle docked and floating as you work.
Type a question, or click a starter prompt to send it as-is.
On Home, Cost, Quality or an agent’s page, open Ask Morse and pick Ask about this page to start from a question about that page.
Watch the step list while Morse works — each tool call it makes shows up as it happens.
From a trace’s Run summary, click Explain this run to get an orchestration-pattern read without typing a question.

⚠️

Morse is hardened against prompt injection: an input scanner and an organization-wide, plan-aware request limit run before your message reaches the model, and an output classifier plus a canary token check for prompt leakage on the way back. A blocked message returns a generic error, not a detailed reason.

Before you have data

With zero traces in your selected workspace, Ask Morse helps you get started with product documentation. Once traces arrive, it can use the data reads enabled for your current Beta scope and access. A first trace does not automatically enable every analysis operation.

To get your first trace here, follow the quickstart.

Beta scope

Ask Morse is in beta, and the beta is deliberately narrower than the full tool set. The panel shows which question types it currently answers, and suggests starter questions from that list. A question outside the list gets a short reply saying so, with a link to the dashboard page that answers it. It reads your data; it does not change anything, so a request to change a setting, delete data, or take an action is outside what it does.

Treat its numbers as a starting point and confirm important figures on the dashboard before acting. The default enables all twelve implemented question categories. A deployment can narrow that list; the panel uses its current list for starter questions. Tenant, workspace, role, subscription and read-only checks still apply to every read. Recognizing a feature does not grant access to it.

Configuration and standalone records, such as alerts, investigations and agent metadata, can be inspected before the first trace arrives. Log search currently reads logs linked to authorized traces. Reads that require traces still point you to onboarding when no traces are present. If my answer service is temporarily unavailable, I explain that instead of treating a valid question as unsupported or asking you to rephrase it.

Privacy and data handling

What happens to a question you ask:

What is sent to the model provider. Your question, the recent conversation, and the results of the tools Morse runs to answer it are sent to the configured model provider. By default that is OpenAI. The provider and model are set by Morse’s deployment configuration, not by you, and Anthropic is a supported alternative.
What the tool results contain. Most tools return aggregates and identifiers: totals, rates, counts, trace and agent names, and statuses. Some return more detail. Trace search and trace detail include a trace’s error message, and the span-detail tool returns a span’s attributes, events, and recorded input and output data. Log search returns the first 500 characters of matching log messages. If a question makes Morse look at a specific span, that span’s content can be sent to the provider.
The same visibility boundary. Supported trace and span reads use the same persisted projections as the authorized UI/API. A recorded user_id, span attribute, or contact-valued field or retained feedback that this user can already see may be returned with an evidence link. Contact-shaped values are not automatically prohibited. Ingestion redaction, payload shedding, access rules, and retention still apply; Ask Morse cannot recover removed originals, enrich an opaque ID into a person’s contact details, or certify what redaction missed. See Security and privacy.
Product visibility and chat retrieval are different. Some product fields are not available through the chat’s current reads. Ask Morse explains that retrieval gap and points to the appropriate view; it does not claim the product never stored the field. Large results are capped, so an omitted field is not proof of absence.
Scoped to your workspace. Tools only read data for your organization and the workspaces your session is scoped to. Answers identify the selected workspace. A link to a broader product view is labeled as navigation; it does not prove that workspace calculation.
Observed data and unsupported outcomes. Recorded metrics describe executed runs. They do not establish the exact cost or quality of an execution that never happened, a future outcome, or completeness of uncaptured data. Ask Morse explains the measurement gap. If it cannot confidently identify a supported read, it asks for clarification rather than treating uncertainty as proof that Morse lacks a feature.
Training. Morse does not use customer telemetry, prompts, or completions to train models. See Security and privacy.

What is kept

Audit log. Each interaction writes an audit record with a SHA-256 hash of your message, the names of the tools used, a hash of the tool arguments, the length of the reply, the model name, and token usage across routing, FAQ verification, and answer calls when the provider supplies it. A provider timeout with missing usage is marked unknown rather than zero. The message text and the reply text are not stored in it. Records older than 365 days are deleted automatically by a daily job.
Conversation state. The conversation is checkpointed in Morse’s database so a thread can continue. The checkpoint holds the messages and tool results of that thread. A thread with no activity for 30 days is deleted, including its checkpoints. A new thread starts after 24 hours of inactivity in a session.
Long-running questions. A question that needs a long window or a wide scan finishes in the background. Morse keeps the question text and the answer for it, with the same retention as conversation threads: 30 days, then deleted by the daily job. The in-app notification and the email carry a link only, never the answer or any figures.
Your trace data follows the retention of your plan, as described in Security and privacy.

Turning Ask Morse off

An organization owner or admin can turn Ask Morse off for the whole organization under Settings, Workspace (the Ask Morse switch). While it is off, nobody in the organization can send a message, the panel shows that Ask Morse is turned off, and no messages are counted against your plan. Other members can see the switch but not change it. Turning it back on restores access; conversations already past the retention periods above stay deleted.

Traces — “Explain cohort” there sends a cohort summary into this same panel.
Trace detail — where the per-trace Explain button lives.