Every fact with a source and a date
[n] citations to what was actually read, server-side removal of unread references, "may be outdated" marks and conflicting sources shown side by side.
Lumen answers from the live data of four products, as the person asking, with numbered sources. It changes nothing without a click, and on Orbit’s cloud it answers with Claude Haiku 5.5 from Anthropic. How many questions an organization can ask is set by its plan.
Orbit Lumen
Lumen reads Orbit, Forge, Pulse and Folio through their APIs as the person asking, so permissions always hold. Every statement carries a numbered source with product, date and author, sources older than a threshold are marked as possibly outdated, and when nothing supports an answer Lumen says so. Changes are proposed as cards that the person runs in their own browser, because the Lumen server has no write path. "Find everywhere" works with no model at all.
Why it is better
[n] citations to what was actually read, server-side removal of unread references, "may be outdated" marks and conflicting sources shown side by side.
No separate index. Lumen reads through each product’s API as you, and a two-user test proves someone outside a project gets nothing from it.
Changes are cards run by your browser. The Lumen server has no write path to any product.
How many questions an organization can ask is set by its plan, and warnings come at 80% and 100%. Find everywhere works without any model.
Tool results travel as untrusted data, with no external images or HTML, and an attack set runs on every delivery.
JQL, Pulse queries, automation rules, CI files and monitors pass the product’s own validator before they are shown or run.
Capabilities
Everything below is listed with its real status. Items with no badge are built and working today.
Built, no badgeIn development Built, in final testingPlanned Designed, not started
A button and Ctrl+J in Orbit, Forge, Pulse and Folio. Streamed answers you can stop.
Statements carry [n] sources. Links to anything the model did not read are removed by the server.
Each source shows product, date and author. Old Folio pages are marked "may be outdated". Conflicting sources show both dates.
No fabricated facts. A hint where to look, checked by an evaluation set.
Continue, rename and delete. Visible only to the author, with retention of 0 to 365 days.
Thumbs up or down with a reason. Post an answer as a work item comment after confirming.
Orbit, Forge, Pulse and Folio data read through each product’s API.
Parallel search over all products with each product’s permissions. Partial results are marked when a product is down.
Orbit validates the query, errors go back to the model and you see the final JQL before applying it.
Validated by Pulse, fixed by error position and opened in Logs.
Someone without access to a project gets none of its data in answers, sources or search. A two-user integration test checks it.
The current entity and selection are published by every section, with suggested questions per entity.
Work items with discussion, Folio pages, merge requests with line comments proposed as cards, failed pipelines with the cause, incidents.
Draft a description, improve or translate selected text and insert into the Folio editor.
Incident, deployment, commits, related Orbit work and the Folio runbook in one cited answer. "Draft a status update".
Lumen proposes. You decide.
Product, fields and preview. Run, edit or dismiss. Nothing changes without a click.
It calls the product API with your session. The Lumen server has no write path.
Create work honouring required fields, comment, transition, link, field updates refused on stale versions and approval requests.
Sub-tasks from a description or epic, editable before creation, created after one confirmation with a report per row.
Automation rules, .gitlab-ci.yml (through CI lint) and Pulse monitors are validated before the card can run.
What was proposed, confirmed or dismissed and the result, visible to the author and administrators.
On Orbit’s cloud, Lumen answers with Claude Haiku 5.5 from Anthropic, with tool calling and streaming. How many questions an organization can ask is set by its plan.
For self-hosted installations: an open-weight model through an OpenAI-compatible endpoint, or a paid provider chosen per function with effort levels, only when an administrator adds a key and a budget ($0 by default). The key is stored encrypted (AES-256-GCM) and shown masked.
Tokens and cost are recorded for every request. On a customer-run installation, the monthly paid budget is $0 by default and stops paid requests with a clear message.
When no model is available the panel says what is missing, and Find everywhere still works with no model.
Product content goes to the model apart from instructions. An instruction found in data never runs without a card.
Markdown without HTML, no external images, external links marked.
Injections in work items, pages, logs and commits. 9 of 9 cases passed twice with no leak and no unconfirmed change.
Only the question, page context and manifest fields. No session, email or user id.
Questions and tool calls limited per person, and the origin is checked on every write.
Providers and models, effort per function, enable per product and project, budgets and limits, retention.
Run reference questions from the console on any model and compare the last 30 runs.
Calls, tokens and cost by model, function, person and agent. The live data check is ahead.
Administrator-only tools over roles and work item security.
Workflow, screens and pipeline explanations with a "what changes after publication" preview.
Name, instructions, allowed tools, projects, visibility and versions, with a test console before you publish.
A product publishes a manifest and Lumen reads it. Orbit’s products use built-in manifests that are validated; joining a new product is a Lumen setting that Orbit controls, so it is not available on Orbit’s cloud yet.
Issue triage, release notes, code reviewer, pipeline doctor, incident analyst, doc writer and administrator helper.
Run on new work items, incidents, triggered monitors, opened merge requests and failed pipelines with the owner’s read-only grant.
Read-only tools for Claude Code and Codex with a per-person token limited to chosen projects.
Compatibility and migration
On Orbit’s cloud, Lumen answers with Claude Haiku 5.5 from Anthropic. Answers still use only what the person asking may see, cite sources, and change nothing without a confirm card.
Answers come from Claude Haiku 5.5 with tool calling and streaming. The number of questions is set by the plan.
For self-hosted installations: an open-weight model through an OpenAI-compatible endpoint, or a paid provider only when an administrator adds a key and a budget ($0 by default).
Read-only tools for Claude Code and Codex with a per-person token limited to chosen projects.
Works with the rest of Orbit
What is next for Lumen
The Anthropic provider that answers on Orbit’s cloud, tested live with plan limits.
An open-weight model through an OpenAI-compatible endpoint, with a paid provider only when an administrator adds a key and a budget ($0 by default). Needs the self-hosted licence.
Calls, tokens and cost by model, function, person and agent.
Why a person has no access, and what a setting will change before you publish it.
Pricing
Lumen is on every plan, with Find everywhere included. Questions are limited per plan.
Orbit is in private preview. Every new organization will start with a 14-day Business trial, no card required.