Your files. Your machine. Your audit trail.

A BOM matching session ends with a single .xlsx workbook. Four sheets, in this order:

Summary — coverage at a glance: matched %, proxy %, gap %, totals by material category. Results — the main table; one row per BOM line, one column group per database, dataset cells hyperlinked to the source record on hiqlcd.com. Gaps — the unmatched and proxy rows, with the databases searched and the next action suggested. Methodology — run parameters and the fixed notes that let a reviewer reproduce the work (system model, GWP100 per IPCC AR6, proxy definition, the practitioner-review requirement).

For the field-level walkthrough — every column, every status icon, every Δ% rule — read /solutions/bom-matching. This page tells you Cowork can produce the workbook. That page tells you exactly what is in it.

Every Cowork project carries a wiki: a local database of structured entries — typed as person, organization, dataset, artifact, concept, event, and others — connected by typed relationships (uses, cites, part_of, created_by, describes, contradicts, supersedes, and more). It is a graph the practitioner can query.

Two views ship in the desktop UI. List shows every entry with search and filter — open one, edit fields in the modal, save. Graph renders the same data as a force-directed graph: nodes float by type-color, edges show which materials share suppliers, which datasets keep getting picked for which rows, which decisions cite which sources. Drag a node to inspect its neighborhood.

A working example. Open Hangzhou Metal Co. — an organization node. created_by edges point to 304 stainless steel coil and 316L, the artifacts they supplied. Each artifact cites an ecoinvent dataset it was matched against, and describes a concept entry that records the APOS-over-cut-off decision the practitioner made for that row. When the same supplier shows up for a new project next month, the graph already knows who and what.

The wiki sits next to the other two project-memory shapes and does what they cannot. progress.md (L5) is an append-only timeline — it answers when did we decide what. The project-memory files (L6, Cortex.md and friends) carry what this project has learned — verified datasets, recurring proxies, supplier specifics. The wiki is structured — it answers how do the things this project knows relate to each other. Audit goes to L5. Convention lookup goes to L6. Knowledge traversal goes to the wiki.

Cowork runs workflows. Schedules run the same workflows on a clock. Daily, weekly, or a full cron expression — Cowork triggers the run, records the result, and posts the next step into the project's timeline.

Three shapes, in increasing project-attachment:

Cortex Dispatch is the phone-side companion to desktop Cowork. Open it on a phone — it installs as a Progressive Web App — and the screen is one text box and one microphone button. Type the task, or hold to record. Dispatch routes it to the desktop Cowork instance signed into the same account; Cowork picks it up and runs.

Voice goes through Cloudflare R2: the phone uploads the audio directly, the server transcribes, and the transcript reaches the Cowork session as the next user message. The Cowork window does not need to be in the foreground. The desktop machine does need to be running and signed in.

A common shape: the procurement manager hears a supplier name in a meeting and dictates "pull the 304 stainless rows from the last BOM and re-match against HiQLCD 1.4." Forty minutes later, back at the desk, the matched workbook is open in the Results sheet and the decisions are written to progress.md. Dispatch did not replace Cowork. It moved the trigger out of the office.

Dispatch is not a separate agent. It is a remote control. Every guardrail Cowork enforces locally — the out-of-workspace gate on file writes, the practitioner-decides pauses on coverage and proxy choice — fires the same way whether the run started from the keyboard or the phone. SSO authenticates Dispatch the same way it authenticates the desktop.

This is what makes Cowork an agent, not a script. It can drive any application installed on your machine — including software with no API and no plugin ecosystem.

When Computer Use is enabled, Cowork gains 21 MCP tools for direct GUI interaction: screenshot, zoom, click (left, right, double, triple, middle), type, key press, scroll, drag, application launch, clipboard read/write, and others. The agent can drive any installed desktop application — openLCA dialogs, a legacy ERP, an internal cost spreadsheet, any app on the machine — without requiring a command-line interface. Before each action, the user approves which applications the agent may control. The main window docks to a side panel; a visible overlay displays "Claude is using your computer." Pressing Escape interrupts at any point.

Computer Use is not available on Windows.

Cowork ships with bundled skills, pulls more from a reviewed marketplace, and writes new ones from your own repeated work. The platform's LCA intelligence accumulates from practice — not from the vendor.

A workflow you've run twice becomes a skill the whole team can reuse.

How skills work — the full story →

Four LCA catalogs are downloaded to local storage at startup — HiQLCD 1.4, Ecoinvent 3.12, CarbonMinds 2.0, Agri-footprint 7.0 — and kept current via verified sync. Catalog search runs locally against the downloaded files. No network call is made for a catalog query. Once downloaded, catalogs work offline.

The out-of-workspace gate runs on every file write and shell command. Any operation whose path falls outside the project folder is intercepted before it executes: the user sees the specific path, a one-sentence reason, and three options — Allow once / Always allow this directory / Deny. The gate fires regardless of session mode. The agent cannot bypass it.

The full data-handling story — what's stored, what isn't, and where — is on the security page.