# Celest

> An entire company behind every person.

Software gave every person tools. AI can give every person a company. Celest is building the operating system for AI coworkers: humans set direction, teammates carry real work, and one shared operating model keeps the company understandable.

Independent software. Not affiliated with or endorsed by Microsoft.

## Reading and service discovery

Public content pages accept `Accept: text/markdown` for their reviewed Markdown representation. Ordinary browser requests return HTML. The same Markdown is also available at the sibling link advertised in each page's HTTP `Link` header.

- [AI Catalog](https://celest.dev/.well-known/ai-catalog.json): One document with connection cards for the corpus-wide Learn MCP and six article scopes. These search published guidance; they do not access live tenant data.
- [API Catalog](https://celest.dev/.well-known/api-catalog): Machine-readable links to those endpoints, their Server Cards, and human documentation.

## Canonical facts

- Company: Celest.
- Category: The operating system for AI coworkers.
- Homepage promise: An entire company behind every person.
- Current commercial starting point: A fixed-scope, read-only Microsoft Agent Estate Review that turns scattered agent evidence into an accountable operating picture and a 30/60/90-day plan.
- Day One promise: Show us your website. We'll build your team.
- Product relationship: Agent Estate Review is the current starting point. Day One is the public proof of the same operating discipline applied to one new assignment.
- Scope: Microsoft is where Celest goes deepest today, not where the company vision ends.
- Team: Scout researches and cites; Manny builds a briefing; Harold independently reviews and accepts or materially revises it. They are generated roles, not Microsoft or Entra identities.
- Personalized context: A submitted public company website seeds cited public-web research; it is not tenant evidence.
- Live demo: When active, Day One runs bounded model and tool work in a disposable workspace with no Microsoft, customer-estate, mailbox, publication, or other external effect.
- Northstar journey: The sample company path uses the same three role-separated turns and runtime as a personalized assignment.
- Recorded/offline evidence: Recorded evidence shows the interaction; it is not a live run.
- Retention: Disposable Day One sessions expire and are deleted within 30 minutes.
- Lifecycle: Context, assignment, roles, capabilities, sources, handoffs, review, authority boundaries, artifact revisions, readback, receipt, reconciliation, and retirement/disposal remain one connected operating trace. Governance is the view that lets you inspect those bindings.
- Current lifecycle: Agent Estate Review discovers and diagnoses read-only evidence. Day One demonstrates intent, team formation, equipping, work, review, and proof when its deployment is active. Connected teams that carry operating context forward are roadmap.
- Authority model: Approval, execution, verification, and reconciliation remain separate gates, so every consequential step has a named owner and proof.
- Trust model: Evidence before action, unknown remains unknown, authority stays explicit, and receipts close the loop.
- Microsoft relationship: Celest complements Microsoft-native platforms and products; it is not affiliated with or endorsed by Microsoft.

## Primary pages

- [Agent Workbench](https://celest.dev/agent-workbench): Celest’s tools for building, evaluating, and repairing Microsoft 365 Copilot agents. The launch page shows the package editing workflow with a synthetic Weekly Brief example, the Make It Yours film, and a Sign up path to the free Celest companion account, with the Microsoft 365 preview guide for host-specific context. [Package guide](https://celest.dev/agent-workbench/docs), [support](https://celest.dev/agent-workbench/support), [privacy notice](https://celest.dev/agent-workbench/privacy) (version 2026-09-11.2, covering hosted package processing, MCP configuration checks, evaluations, and companion accounts), and [terms guide](https://celest.dev/agent-workbench/terms) (which agreement applies by channel). Also available as [Markdown](https://celest.dev/agent-workbench.md).

- [Celest](https://celest.dev/): One place to understand and operate agents across Microsoft. The homepage leads with Agent Estate Review and uses Day One as public proof of the larger company thesis. Also available as [Markdown](https://celest.dev/index.md).
- [Agent Estate Review](https://celest.dev/agent-estate-review): A read-only review that connects fragmented Microsoft agent records into one accountable operating picture and a 30/60/90-day plan. Also available as [Markdown](https://celest.dev/agent-estate-review.md) and [JSON](https://celest.dev/agent-estate-review.json).
- [Estate Review preview](https://demo.celest.dev/reviews): Launch a preview of the Estate Review mission at `/reviews`.
- [Day One team experience](https://demo.celest.dev/): Bring a public company website or start with Northstar, then watch the team research, build, and review. Live work and recorded playback are labeled directly.
- [Sample Estate Review readout](https://celest.dev/agent-estate-review/sample): The shape of the deliverable on a labeled illustrative estate — canonical agent records, evidence-backed findings, named unknowns, a 30/60/90-day plan, and source coverage. Also available as [Markdown](https://celest.dev/agent-estate-review/sample.md) and [JSON](https://celest.dev/agent-estate-review/sample.json).
- [The Agent Estate Review methodology](https://celest.dev/learn/agent-estate-review): Source-backed five-pass method for bounding an estate, qualifying evidence, preserving unknowns, producing deterministic findings, and creating an operating plan. Also available as [Markdown](https://celest.dev/learn/agent-estate-review.md), [JSON](https://celest.dev/learn/agent-estate-review.json), and a [Scoped MCP](https://celest.dev/learn/agent-estate-review/mcp).
- [Celest Learn](https://celest.dev/learn/): Canonical, source-backed answers about operating and governing AI agent estates.
- [Celest blog](https://celest.dev/blog): Dated implementation notes and arguments from building Celest, including [We tried converting Agent Plugins to Microsoft 365 Cowork.](https://celest.dev/blog/we-tried-converting-agent-plugins-to-microsoft-365-cowork)
- [What is an AI agent control plane?](https://celest.dev/learn/ai-agent-control-plane): Celest's definition, seven-stage operating loop, minimum canonical record, authority model, and limitations. Also available as [Markdown](https://celest.dev/learn/ai-agent-control-plane.md), [JSON](https://celest.dev/learn/ai-agent-control-plane.json), and a [Scoped MCP](https://celest.dev/learn/ai-agent-control-plane/mcp).
- [Evidence, authority, and receipts](https://celest.dev/learn/evidence-authority-and-receipts): Celest's control model for keeping evidence, findings, proposals, authority decisions, handoffs, execution observations, and terminal receipts distinct. Also available as [Markdown](https://celest.dev/learn/evidence-authority-and-receipts.md), [JSON](https://celest.dev/learn/evidence-authority-and-receipts.json), and a [Scoped MCP](https://celest.dev/learn/evidence-authority-and-receipts/mcp).
- [Analytics is evidence, not authority](https://celest.dev/learn/analytics-is-evidence-not-authority): A practical guide to using adoption, cost, and value analytics as evidence without mistaking a dashboard for ownership, authority, execution, or verified state. Also available as [Markdown](https://celest.dev/learn/analytics-is-evidence-not-authority.md), [JSON](https://celest.dev/learn/analytics-is-evidence-not-authority.json), and a [Scoped MCP](https://celest.dev/learn/analytics-is-evidence-not-authority/mcp).
- [A practical guide to governing AI agents at work](https://celest.dev/learn/agent-governance-controls-matrix): A friendly guide to finding your AI agents, assigning owners, checking access and activity, and following changes through to verification. Also available as [Markdown](https://celest.dev/learn/agent-governance-controls-matrix.md), [JSON](https://celest.dev/learn/agent-governance-controls-matrix.json), and a [Scoped MCP](https://celest.dev/learn/agent-governance-controls-matrix/mcp).
- [How to govern AI agents across Microsoft](https://celest.dev/learn/microsoft/agent-governance): A practical guide to reviewing one agent across truthful but partial records in Microsoft 365, Agent 365, Entra, Copilot Studio, Power Platform, Purview, Defender, Foundry, and GitHub. Also available as [Markdown](https://celest.dev/learn/microsoft/agent-governance.md), [JSON](https://celest.dev/learn/microsoft/agent-governance.json), and a [Scoped MCP](https://celest.dev/learn/microsoft/agent-governance/mcp).
- [Privacy](https://celest.dev/privacy): Celest's three scoped privacy boundaries: first-party, no-cookie website analytics; covered no-Celest-runtime package-only Microsoft 365 agents; and support communications about the site, notice, or covered packages. A Celest listing or immutable release record must expressly opt a package in; Helpdesk Triage uses a product-specific notice. Includes a [machine-readable event manifest](https://celest.dev/analytics/manifest.json) and [Markdown](https://celest.dev/privacy.md).
- [Terms of Service](https://celest.dev/terms): Celest's free/included B2B package-only Microsoft 365 agent and plugin baseline, built as a Common Paper Cloud Service Agreement Cover Page incorporating Common Paper Standard Terms v2.1. Also available as [Markdown](https://celest.dev/terms.md).
- [Helpdesk Triage](https://celest.dev/helpdesk-triage): Limited preview — enrolling by conversation. AI ticket triage for SharePoint service desks: a tenant-wide SharePoint Framework (SPFx) package with a Copilot Component, a page web part and a list command, backed by SharePoint lists you own. For the tickets that stall because nobody can say which team owns them, Microsoft 365 Copilot proposes one exact change as tool arguments with the evidence marked observed or inferred and what is still missing named; the signed-in operator edits and approves in the card; one ETag-protected SharePoint patch is applied and a fresh readback records verified, stale, diverged, indeterminate or failed on the run row. The queue appears in Microsoft 365 Copilot, on a SharePoint page and from the tickets list's command bar; proposals and approvals happen in Copilot and from the command bar. Approvals happen in the Copilot card or from the tickets list's command bar; the first release does not deliver approvals through Teams or Agent 365. Ticket data stays in your SharePoint lists; Celest runs no service and receives no ticket data. By 2026-09-04, in Celest's own tenant, the component loaded in SharePoint's Copilot Workbench had taken ticket HD-07 from a proposal fired by hand to an operator approval to a verified SharePoint readback receipt, and the shipped package had verified a hand triage from the tickets list's command bar. On 2026-09-04, inside Microsoft 365 Copilot, Copilot wrote a proposal on the sandbox's sample data and an operator's approval verified. The shipped package's own agent has not yet run inside Microsoft 365 Copilot chat. Free during the preview. Start a limited-preview conversation through support@celest.dev. Independent software. Not affiliated with or endorsed by Microsoft. Also available as [Markdown](https://celest.dev/helpdesk-triage.md).
- [Helpdesk Triage limited preview](https://celest.dev/helpdesk-triage/preview): Limited preview — enrolling by conversation. Bring one SharePoint queue and the ticket that has stalled longest; a founder walks it from queue to proposal to approval to receipt with you, using your routing vocabulary, roles, and review practice, through support@celest.dev. Free during the preview. By 2026-09-04, in Celest's own tenant, the component loaded in SharePoint's Copilot Workbench had taken ticket HD-07 from a proposal fired by hand to an operator approval to a verified SharePoint readback receipt, and the shipped package had verified a hand triage from the tickets list's command bar. On 2026-09-04, inside Microsoft 365 Copilot, Copilot wrote a proposal on the sandbox's sample data and an operator's approval verified. The shipped package's own agent has not yet run inside Microsoft 365 Copilot chat. Also available as [Markdown](https://celest.dev/helpdesk-triage/preview.md).
- [Try Helpdesk Triage in your browser](https://celest.dev/helpdesk-triage/try): Limited preview — enrolling by conversation. The cards the product ships, on eight sample tickets kept in your browser's own storage; no sign-in and no tenant, and Reset clears it. Open a ticket, ask for a scripted proposal, change a field, approve, and read the receipt. The sandbox scripts the proposal so you can see the loop without a model; in the product, Microsoft 365 Copilot writes it from the queue it is shown. The sandbox server answers the page's tool calls, stores no sample data and sets no cookies; Cloudflare keeps ordinary request logs for a few days. Also available as [Markdown](https://celest.dev/helpdesk-triage/try.md).
- [Helpdesk Triage privacy notice](https://celest.dev/helpdesk-triage/privacy): The package runs in your Microsoft 365 tenant as the signed-in user and keeps everything in your SharePoint lists; Celest receives only support mail and, if the package is acquired through Microsoft's marketplace, Microsoft's acquisition report. Lists what setup creates in the tenant (three lists, a Default routing item, optional sample tickets, the site agent file, the list command, the main-agent setting) and how to remove it, retention (your SharePoint settings), providers (Microsoft in your tenant; Cloudflare for the website and the sandbox; Celest's own Microsoft 365 tenant for support mail), the sandbox, and contact. Notice version 2026-09-04.2. Also available as [Markdown](https://celest.dev/helpdesk-triage/privacy.md).
- [Helpdesk Triage docs](https://celest.dev/helpdesk-triage/docs): The operating contract for the package: the first-release boundary (a tenant-wide SPFx package with a Copilot Component, a page web part and a list command), the data boundary (everything stays in your SharePoint lists; nothing is sent to Celest), the loop (queue, proposal, approval, one update, readback), the receipt contract, the real statuses (New, In Progress, Resolved, Awaiting triage, Triaged) and receipts (verified, stale, diverged, indeterminate, failed), product limits, the first-release boundary (Approvals happen in the Copilot card or from the tickets list's command bar; the first release does not deliver approvals through Teams or Agent 365.), and pricing: Free during the preview. The Microsoft marketplace listing is prepared for SPFx 1.24 general availability; until then the package installs from your own tenant app catalog, and nothing is charged for the preview. Also available as [Markdown](https://celest.dev/helpdesk-triage/docs.md).
- [Helpdesk Triage product terms](https://celest.dev/helpdesk-triage/terms): Product terms for the free Celest Helpdesk Triage package, which incorporate the Celest Terms of Service (Terms Version 2026-09-03.1) by reference with a product-specific definition: client-side code in your tenant, no Celest runtime, proposals applied only after a person approves, no fees, best-effort support, removal at any time. Product terms version 2026-09-04.2. Also available as [Markdown](https://celest.dev/helpdesk-triage/terms.md).
- [Helpdesk Triage support](https://celest.dev/helpdesk-triage/support): Email support@celest.dev with the minimum context for product and setup questions, incidents, uninstall and deletion instructions, and data-request routing. Support cannot change a ticket; only an operator’s approval in the card does. Also available as [Markdown](https://celest.dev/helpdesk-triage/support.md).
- [Helpdesk Triage setup help](https://celest.dev/helpdesk-triage/support-help): Getting Helpdesk Triage ready for your SharePoint service desk: what to have ready (site and queue, roles, routing configuration, Microsoft 365 Copilot), what the setup card creates on your click, a rehearsal with one ticket, and an email to support@celest.dev with the useful context. Also available as [Markdown](https://celest.dev/helpdesk-triage/support-help.md).
- [Helpdesk Triage security](https://celest.dev/helpdesk-triage/security): Nothing runs outside your tenant: the package runs as the signed-in user with existing SharePoint permissions, holds no credential, calls no Celest endpoint, applies a change only after approval with the ticket’s version check, and reads it back. What it never does, what setup creates, and how to report a concern. Also available as [Markdown](https://celest.dev/helpdesk-triage/security.md).
- [Helpdesk Triage uninstall](https://celest.dev/helpdesk-triage/uninstall): Remove the package from the tenant app catalog, delete the list command, decide about the three lists you own, delete the site agent file and clear the main-agent setting, and remove sample tickets. Celest holds no product data, so there is nothing to ask Celest to delete. Also available as [Markdown](https://celest.dev/helpdesk-triage/uninstall.md).
- [Celest on GitHub](https://github.com/celest-dev): Public open-source work from Celest.

## Agent interface

- [Celest Learn MCP](https://celest.dev/learn/mcp): Public, anonymous Streamable HTTP server with read-only retrieval and optional feedback submission. `resources/list` exposes the curated Celest reading list; `ask` uses Exa discovery and external source content, with up to two owned Celest publication candidates added for unscoped questions, followed by a tool-free relevance selection pass and returns up to two sources by default, at most six, with up to 12,000 characters and 20 useful links per source. The truncation marker appears only when source text reaches the 12,000-character bound; the link list is capped separately.
- Every canonical Learn resource advertises a scoped `/mcp` child with the same retrieval and feedback tools. Scoped `ask` stays within that article; scoped `similar` traverses to neighboring published Celest resources.
- A useful first question is `What is an Agent Estate Review, and what evidence does it collect?` or `Where should I look across Microsoft to review one agent's ownership and activity?` Use `resources/list` to inspect the curated reading list, `ask` once for cited excerpts, then open a cited source URL in a browser for full context. Use `similar` for the next Celest resource and `resources/read` for the exact full Celest Markdown publication.
- The optional `feedback` tool accepts a result ID, usefulness rating, optional reason, and short note. After the task, use it only for a concrete first-hand observation such as a directly resolved question, missing detail, or broken source; a generic thank-you alone is not evidence. Follow host permission rules, prefer the result ID and outcome, add a note only when needed, avoid repeated solicitation, and leave out private conversation content, personal information, and credentials.
- [Improve Celest Learn](https://github.com/celest-dev/learn/issues/new?template=learn-feedback.yml): report a correction, missing topic, source suggestion, or retrieval result. Include the canonical URL, complete question, and public passage or source when available; a source digest is optional. Never include credentials or private tenant data. If GitHub is inconvenient, email founders@celest.dev.

## Evidence levels

- Live: a mature capability Celest commits to contract, deliver, operate, and stand behind for every admitted customer scope.
- Live when active: the Day One disposable model/tool journey described by the exact reviewed deployment and its receipt; it is not a Microsoft/customer effect.
- Sample: a named, fictional company journey that still runs through the same real three-turn runtime when the live deployment is active.
- Recorded/offline: deterministic browser evidence that demonstrates a projection without claiming a contemporaneous model or provider run.
- Experimental: proven research behind replaceable or private contracts; not part of the durable public API.
- Roadmap: intended product direction, not shipped capability.

## Contact

- Email: founders@celest.dev