What project holds
A project’s navigation lists Home, Chats, Workflows, Knowledge, Context, Outputs, Connections, and Settings.
Where projects live
A project lives in one workspace. Everyone in the workspace sees it, and their role in that workspace decides what they can do in it. A person can also hold the Builder role on the project itself: a project builder can create, link, and relate artifacts in the project. Editing an existing Claw or schedule is the exception: it follows what that resource’s own permissions allow. See Roles and access for the roles. In the app rail, select Projects to open the list. It follows the scope you have open: every workspace’s projects when the scope is the whole account, and one workspace’s projects when you have opened that workspace. Each card shows the project’s name, workspace, goal, sources, and when it was last updated, most recent activity first. The list separates Active and Archived. From an open project, the project’s mark and name at the top of its sidebar are a switcher: they open a searchable menu of the active projects in workspaces you have joined — the ones you opened most recently on this device first — and picking one opens its Home, so you move between projects without going back to the list.Create project
- In the app rail, select Projects, then New project. The button appears when you can create a project in a workspace you can reach.
- Write what the project is for in a sentence. The sentence becomes the project’s Goal, which agents read first in every session. Once you start typing, the app proposes a Name from the sentence, which you can change; under it is the handle the project’s link and the CLI use, which can’t change later.
- Pick the workspace the project lives in from the picker under the sentence. It offers the workspaces you can start a project in, and everyone in the workspace will see the project.
- Leave Add Claw on to create the project with its agent: it learns what the project is about, suggests where to start, and helps with the work using your access. If you aren’t an admin of the chosen workspace, the switch is locked and reads Claw needs a workspace admin: the project works without Claw, and an admin can add it later.
- Select the create button. Its label names the project, such as Create Revenue data, or reads Create project when the name runs long. The project opens.
Add to project
A project opens on its Home. To add a skill, open Context and select + beside Skills. Write a skill writes a new one in the project; Add from your organization links one your team already uses. The picker lists the project’s own workspace’s artifacts and marks what the project already has as Already in this project. To connect a tool or a key, open Connections and select +: MCP server takes the server’s URL, API or key takes a base URL, a key, or a connection string, and From your workspace adds one your workspace already has. The MCP servers in that list also include the organization’s servers that are available in the project’s workspace, each marked From your organization. Other kinds, such as agents, have a tab of their own in the project: it lists the project’s items of that kind with their Origin — Created here when the artifact was created in the project, Linked when it was linked in from the workspace — and when they were last Updated. If you can edit the project, the tab’s Link button (such as Link agents) opens a picker that links artifacts one kind at a time. To remove an item, open its actions menu, choose Remove from project, and confirm. On Connections, a connection that holds more than one artifact names the kind instead, such as Remove the API from project. The item leaves the project’s contents and stays in its workspace. Nothing is deleted, and access is unchanged. You change a linked artifact where it lives; the project shows the change.Set up the project’s agent
A project comes with its own agent, Claw. Claw is named after the project: the project’s name plus Claw, so a project named Revenue data has a Claw named Revenue data Claw — the name the app lists it under, and the one the consent page asks a person to authorize when they let Claw act as them. When the project’s Claw still needs setting up — the project was created without it, or its setup didn’t finish — the project’s Home shows a setup card for it: Claw reads the workspace, keeps notes you can correct, answers in Slack, and runs your schedules. Claw can discover tools across the MCP gateway, and each tool’s own access rules decide what its calls may do. Claw can also call the external REST APIs your team registered: it lists the project’s APIs withextension__api__list and calls one with extension__api__call, and the gateway injects each API’s linked credential, so Claw never asks anyone for an API key. It can call only the APIs homed in or linked into the project; to use another, ask a person to link it into the project. You can let Claw act as you: the schedules it saves then run as you, with your connections. Your own sign-ins to integrations count as connections too: when Claw runs an ask as you, it reaches the integrations you have signed in to, and the project’s Home card Connected lists each one as Your sign-in — a system the project already connects is not listed twice. When a run needs sign-ins you haven’t connected, Claw asks for them all at once in the chat — under a line such as Two connections need you before Claw continues — and notes where your sign-ins are kept. You can see every tool you’ve signed in to, and disconnect one, under Connections in your account settings.
Select Set up Claw to start. Only an admin of the project’s workspace can give Claw its own identity, so if you are not one, the card says so and offers Copy link for an admin: the project’s link is copied, and an admin who opens it can set Claw up from that page.
While Claw sets up, the card shows its progress, and you can leave the page — setup keeps going. Setup ends when the first read starts, so the read shows as Claw working, not as setup: while it runs, the success notice says Claw is reading and offers Watch it read to follow that session.
Setup also gives Claw a heartbeat: an hourly schedule called Claw heartbeat that starts a run only when the project has changed. Each run reads what changed since the last one and keeps Claw’s memory bundle current — it adds to the day’s note and writes the suggestions the project’s Home shows. A heartbeat run is upkeep, not a chat: it doesn’t appear in the project’s chat list. A run reads at most the last day of changes, so resuming after a long pause does not replay everything since. Pause the heartbeat like any of Claw’s schedules, and it stays off until someone turns it back on. The heartbeat is the platform’s maintenance: you can pause, resume, or archive it, but you can’t run it off its schedule or edit it. Archiving it ends the memory upkeep for good — the confirmation suggests pausing instead when you mean only to stop it for now.
Claw has its own face: it shows what Claw is doing — working, waiting, needing you, or paused — and each project’s Claw wears its own colour, the same for everyone who opens the project. That mascot is the project’s icon — the way the project shows in lists, the sidebar, and chats — unless someone who can edit the project picks an emoji for the icon on Settings. The mascot’s colour stays the same when you rename the project.
Changing Claw once it is set up follows what each thing’s own permissions allow. You rename the Claw when you can edit the Claw itself, and change its instructions and what it may do when you can save a Claw version — those changes become Claw’s next version. Renaming the Claw keeps every person’s consent: someone who let the Claw act as them under its old name doesn’t have to let it again. Renaming the project doesn’t rename its Claw. A Claw still named just Claw — the name every Claw was created with before Claws were named after their project — takes the project’s name the next time it runs; a name a person gave the Claw stays. Its schedules follow their own permissions too: pausing stops only the ones you may edit, and Claw tells you how many it leaves on; archiving a schedule needs that schedule’s archive permission; adding one needs permission to create content in the workspace. An archived project’s schedules cannot be changed.
Claw’s standing instructions live on Context as agent.md, the file Claw reads before every chat and run. Edit saves Claw’s next version, and Claw’s tools beside it change what Claw may use; a Claw with no instructions yet says so there.
Chat with Claw
Ask Claw in the project’s chats, or in the Claw bar on any of the project’s pages. Each chat is one of the project’s sessions, and the project’s chat list shows everyone’s chats and your own private ones. Claw’s replies read as a few short texts rather than one long block: each arrives in its own bubble as Claw writes it, and its typing dots stand in while it works — its steps and reasoning stay out of the chat. It points at things by name: the name in a reply is a link, and selecting it opens that thing in the app — a workflow, a knowledge note, a schedule, a chat, or an output — while a link that opens outside Atlan says so. When what a link named is gone, the name stays as text that opens nothing, and the tooltip says No longer available. Claw can also open the conversation itself. On a project whose Claw nobody has talked to yet — one you just created, for example — Claw starts a chat to get to know you and the project, and speaks first in it. Home shows Claw’s hello in a bubble beside the greeting, leading with what it asks or offers, until you select See Claw’s message to open that chat in the Claw bar or put the bubble away. While the hello is up, the composer on Home becomes Reply to Claw: what you type there opens the chat in the Claw bar with your message sent. The Claw bar’s window also holds a workflow’s draft and its changes: describing a workflow on Workflows starts one in the window over that page, and so does asking for a change on a workflow’s own page. While the draft lasts, the window stands in for the page’s composer. Open leads to the workflow once the draft has one, and New chat ends the draft and gives the page its composer back. Claw reads the project’s public chats, so it can recall an earlier conversation when you ask what was said before. It never reads a private chat or a chat started from Slack. When you ask Claw to remember, correct, or forget something, it completes that change in its memory bundle before it reports the request done. Saving a deliverable — a file, or an output in the Library — does not itself update the memory. If Claw cannot save the memory change, it reports that part of the request as incomplete, alongside any deliverable it saved. By default a chat is public, so everyone who opens the project sees it in the chat list. To keep one out of the list for everyone else, turn on Private chat — the dotted bubble in the chat page’s toolbar, or the same toggle in the Claw bar — before you send your first message. The page’s heading becomes Private chat, and the note under it says what that does: the chat stays out of the chat list for everyone else in the project. Claw cannot read it either. In the chat list, a private chat shows a dotted bubble in place of the author’s face, and the open chat shows Private in its toolbar. You make the choice when the chat starts; New chat in the Claw bar keeps it for the chat that follows. Over the API, the same choice goes with the session you create:POST /agent/v1/agents/{agent_id}/sessions takes is_private.
Run workflows
On Workflows, the work you described becomes a workflow with runs of its own: its page lists the runs it has made. A run replays step by step — what each step did and returned — and a failed step offers Rerun from this step and Ask Claw what happened. A workflow that reads one person’s connection or publishes to Slack is its owner’s to turn on. Its page shows an Activation review first: the exact version the review covers, who it runs on behalf of, the agent that runs it, each source connection and whether it is connected, the schedule and the next run, and — for a publishing step — the Slack sender, the destination, a preview of the attribution line the message carries, and a Delivery readiness warning when the sender cannot reach the destination. A What will be shared note says in words what the workflow will share and through whose connection. Activate turns the workflow on and Deactivate withdraws the owner authorization; while the review names something to settle — connect a source, invite the app to the channel, choose a destination, let the runner act for the owner, or review the version’s changes — its notice offers that one action. Only the workflow’s owner can activate it, and the review’s state stands where an on switch used to: Active, Needs review, Needs attention, or Inactive. Turning the workflow on without its review refuses and points to the review, and pausing a workflow with an authorization in force — including when its project is archived — withdraws that authorization, so turning it on again runs the review again. A workflow with neither a personal connection nor anything to publish still turns on with its switch. You choose a publishing step’s sender and destination in its Slack publishing section, in the draft’s panel or on the step’s detail. The sender is Publish as me — your Slack profile, with a “Sent using Trident” footer — or Publish as Trident, the Trident Slack app bot. The destination is one of the project’s channels, and a personal sender can keep the personal channel already saved on the step. Both a sender and a destination are required, and Save publishing step saves the choice on the step; a destination the sender cannot reach cannot be saved, and when the app bot is not in the channel the note says to invite it there. A failed run says which phase stopped: reading its sources — with a nudge to check the owner’s connections — creating its result, or delivering it; a workflow can also fail on a configuration change it needs, or stop on a service error. A run without a named phase keeps the reason the run recorded. When the result was created but did not arrive, the notice offers Retry delivery to anyone who can run the workflow, and the retry reuses the generated result without reading the sources again. When the last attempt may have posted anyway — the app cannot tell — the notice says to check Slack first, and the retry asks to confirm, because the message may be posted twice. What a run gathered through the owner’s personal connections is hidden from everyone but the owner: where a result would be, the page says only that it was gathered through the owner’s personal connections, and a step that ran on them keeps just the text it delivered.Read the project’s knowledge
On the Knowledge tab, Claw and the project’s workflows keep knowledge bundles: pages written as they work. Open a bundle to read it in place — the bundle’s Pages beside the open one, and the open page rendered with its headings, so links between pages stay inside the bundle. Beside the page, About this page shows what the registry knows about it: Written by, when it last Changed, and, when a page has them, its Version, Status, how far it has been Checked — not reviewed yet, checked by a tool, or reviewed by a person — and its Tags. Select Open in Brain to open the bundle in the Brain. In the Brain, the bundle reads as its pages: the Pages list beside the open one, searchable, with the bundle’s Contents and History at its ends — pages the knowledge app writes unless the bundle holds its own file at that path. A page shows, when it has them, its type and how far it has been checked, and a Past review date warning once its review date has passed. Opening a page the registry has indexed opens its bundle on that page, so the page reads beside the rest of it; one the knowledge app has not filed in a bundle yet shows This page has no bundle. Ask about the page in the prompt under it, and Claw reads that page first, then the rest of the project’s Knowledge. A bundle with no pages yet says so: ask Claw on Knowledge to start one.Sources
Open the project’s Sources page. Paste a web link starting withhttps:// or http:// and add it. A project holds at most 32 sources, and the page counts how many are used. Remove a source from its row.
Change or archive project
On Settings, the project’s Profile holds its Icon, Name, and Goal. Select Change icon to pick the project’s icon: Claw by default, or an emoji in its place. The Name and Goal fields save on their own — on Enter, or when you leave them — with a notice when one saves. The CLI name is shown and never editable: it is how the CLI and agents refer to the project, and it is set when the project is created. The Slack section lists the channels connected to the project: mention Claw in one to start a chat in this project, and replies stay in the thread. If you can edit the project and its Claw is set up, select Connect channel to add a channel Atlan’s Slack app is already in. Disconnect on a channel’s row asks to confirm: mentions in that channel no longer start chats in this project, and its existing threads stay. Danger zone holds Archive project; only a workspace admin can archive a project. When Claw’s schedules or the project’s workflows are still running, the confirmation offers Also pause Claw and the workflows (recommended), on by default. Archiving hides the project from the active list and keeps its history. Nothing linked into it is touched: every skill, agent, and source linked there stays where it lives. Archived projects are on the list’s Archived tab.Example
A platform team starts Q4 pricing launch in their shared workspace. The goal says the project ships the pricing page, the FAQ, and the sales enablement kit. They link the SQL review skill and the data steward agent, and add the pricing model spreadsheet and the draft product brief as sources. A teammate opens the project and starts a session, and the session begins with the linked skills and the goal, so nobody pastes the same links into the prompt.What projects do not do
- A project is not a second home for an artifact. An artifact added from your organization stays in its workspace, and removing it from the project deletes nothing.
- A project grants no access to the artifacts it links. Everyone in its workspace sees it, and their workspace role decides what they can do in it; a project builder can also create, link, and relate artifacts in it. Editing an existing Claw or schedule follows what that resource’s own permissions allow.
- An archived project is not deleted. Archiving hides it from the active list and keeps its history, with its links untouched.
See also
- How workspaces work: The scope a project lives in and inherits.
- How sessions and traces work: What a session records once it runs.
- Choose where skills live: Decide the workspace before you publish.