Overview
What the Telflo AI assistant does and how to use it.
The AI assistant is a chat panel docked at the right edge of the editor. It can plan a pipeline from a description, generate or modify the configuration on the canvas, test the result, and answer OpenTelemetry questions.
Open the assistant
Select the AI button in the editor toolbar. The assistant opens as a docked panel; the navigation menu and the editor's right panel collapse to icon rails to make room, and expand again when you hide the assistant with the AI button or the panel's close control. The panel is resizable by its left edge, keeps the chat while you work on the canvas beside it, and reopens at its last size on your next editor visit.
The AI button appears when the Enable AI Features setting in Settings > AI is on, which it is by default. If your membership in the current organization has not been granted AI access, the panel shows a request prompt instead of the chat. See AI access.
What you can ask
Create a configuration:
- "Create a pipeline that receives OTLP and exports to two backends."
- "Build a Kubernetes monitoring setup with metrics and logs."
Modify the current configuration:
- "Add a batch processor to the traces pipeline."
- "Change the OTLP exporter endpoint to
otel.example.com:4317."
Ask a question:
- "What is the difference between OTLP and OTLP/HTTP?"
- "How does tail sampling work?"
How it works with the canvas
The assistant reads and edits the configuration you are working on, and consults reference material as it works.
| Tool | Action |
|---|---|
| Read current YAML | Reads the current configuration, so answers and edits are based on what you have. |
| Update YAML | Applies a change to the configuration. The canvas and the YAML update to match. |
| Validate YAML | Validates generated configuration against the collector binary before it is applied, up to five attempts per request. $orgvar: references are replaced with the organization's default values in a transient copy, so the configuration keeps its tokens. When the validation service cannot be used, because the validator cannot be reached, does not respond in time, or rejects the request, the check is skipped rather than treated as a failure: the tool reports the reason, the assistant applies the change and tells you the configuration was not binary-validated, and the editor's status row shows validation unavailable · config not checked until your next change is checked. A collector version the validator does not support is still reported as a validation error. |
| Reference lookup | Consults curated OpenTelemetry reference documentation for components, OTTL, and naming conventions. The chat shows a progress line while a reference is consulted. |
| Semantic conventions | Queries a pinned release of the OpenTelemetry semantic-conventions registry for attribute, span, and metric names, in Build and Ask modes. |
| OTTL syntax check | Checks OTTL expressions against the collector's grammar before they are used. |
| Run a pipeline test | Runs the configuration in the testing sandbox against a sample and evaluates assertions on the output, in Build mode, with the same organization-variable substitution as the Testing tab. Up to three runs per message; the runs appear in the test history. |
| Save validation rules | Saves assertions from a test to the pipeline's validation rules, in Build mode. See Validating output. |
| Web search | Looks up current OpenTelemetry information as a fallback, for topics the built-in references do not cover. |
Tool calls appear in the chat as collapsible cards while they run, and update in place when they finish. Expand a card to see the request and the response. Tool cards are saved with the conversation.
When the assistant updates the configuration, the change appears on the canvas as if you made it. Review, edit, undo, and save it as usual.
The assistant validates and runs pipeline tests against the OpenTelemetry Collector Contrib distribution only. For a configuration locked to another distribution, such as ClickStack in private preview, its Validate YAML reports the configuration's collector version as not supported and Run a pipeline test cannot run; after five failed validation attempts, the assistant applies the change anyway and lists the errors. Use the editor's own validation and the Testing tab for those configurations.
Organization variables
The assistant treats $orgvar:NAME and $uservar:NAME tokens as references it must preserve. When it modifies a configuration, it keeps every token exactly as written, never replaces one with a literal value or an ${env:NAME} reference, and adds a new token only when you ask for one. For its own validation and test runs it substitutes the organization's default values into a transient copy, the same way the editor's validation does. If validation fails on a token that refers to an organization variable your organization has not defined, the assistant keeps the token and asks you to define it in the vault rather than inventing a value. In Ask mode, the assistant treats tokens as intentional and does not advise replacing them.
Agents outside the editor
The assistant works on the configuration open in the editor. To let an agent you run elsewhere, such as Claude Code, Cursor, or claude.ai, read, validate, test, and deploy your organization's configurations, connect it to the MCP server. Pipeline tests an agent starts through the MCP server do not substitute organization variables, unlike the assistant's own runs. MCP server, CLI, or AI assistant compares the three.
Modes
Select a mode in the composer to direct the assistant: Plan, Build, or Ask. See Modes.
Conversations
Chats are saved, and Telflo generates a title for each one. The conversation list shows your chats from the last 30 days.
Usage and safety
AI usage is metered with tokens. See AI access. Requests and responses pass through guardrails; which checks run depends on the mode.
Last updated on