AI that actually does things.
Ask Atlas answers questions from your workspace, with citations. It captures tasks from plain language, writes a daily standup digest, summarizes meetings, and plans work. Agents act only through the tools you allow, wait for approval where you require it, and every action lands in the audit log. Ask Atlas is on paid plans.
Four layers, built for action.
The Assistant is not one model. It works in four parts: memory, planner, executor, and audit. Each one is visible to the people who run it.
Memory entries scoped to the workspace, one agent, or one user. Entries can expire, and you can list and delete them.
Turns a goal like "draft the recap and book the follow-up" into steps. Each step names the tool it will use. You can read the plan.
Runs only the tools an agent was given, inside the caller's permissions. Tools marked for approval wait for a person to say yes or no.
Each action is written to the audit log with the tool used and its scope. An approved action can be replayed from its approval.
Five things that compound.
A chatbot summarizes. The Assistant also acts. Acting safely takes identity, permissions, approvals, and audit, so those are built in.
Ask Atlas retrieves context from your tasks, projects, comments, and docs, and returns an answer with citations you can check.
The caller's role and module permissions apply exactly as they do in Atlas. Each agent can only use the tools it was given.
You choose which tools need approval, such as external sends or contract steps. The held action waits in a queue until someone approves or rejects it.
Memory entries are scoped to the workspace, an agent, or a user. They can expire on a date, and you can delete any of them.
Agents can run on a schedule with a goal and a monthly cost cap, and can start from agent templates. The daily standup digest is built to run every morning.
Three phases, per action.
A user asks. The planner turns the ask into steps. The agent runs them in order, with tool scope, approvals, and audit on each. Three phases, one log.
Decompose to typed steps
The planner produces a plan you can read before it runs. Each step names the module and the tool it will use.
Gate every side effect
A step runs only if the agent has that tool and the caller has the permission. Tools marked for approval hold until a person decides.
Log every event
Each action writes an audit log entry with the user, the tool, and its scope. Approved actions can be replayed, and admins review it all in one place.
Nine more, all in policy.
Ask Atlas and agents are on paid plans. Each of these comes with them.
Per-user identity
Actions run with the caller's role and permissions, and the audit log records who started them.
Approval tools
Mark any tool as approval-required per agent. Pending approvals sit in one queue.
Memory you can read
List memory entries, see their scope and expiry, and delete any entry you no longer want.
Agent templates
Start from a template, then adjust the instructions, tools, and approvals to fit.
Daily standup digest
A short summary of yesterday's activity and today's plan, generated per person.
Your model, your keys
Use an outside AI provider with your own keys, or point Atlas at your own model server.
Encrypted in transit and at rest
TLS 1.3 on every request. Data is encrypted at rest by our managed hosting provider.
No training on your data
Your content is not used to train models. Private AI runs its own models on Atlas servers.
MCP server
Use Atlas from Claude Desktop, Claude Code, Cursor, or any MCP client, with tool scopes.
Works across your modules.
Ask Atlas and agents work across your Atlas modules, with the same permissions as the person who asked. Nothing happens behind anyone's back.
One assistant for the whole workspace
Agents act through named tools, and each tool has a scope. The caller's role and module permissions apply exactly as they do in Atlas, so permission boundaries hold, and each action is written to the audit log.
Programmatic Assistant.
Ask questions, run agents, and approve held actions over the REST API. Calls need the right scopes, and role permissions apply.
# Ask Atlas: a grounded answer with citations. curl -X POST https://api-atlas.wrxstack.com/v1/ai/ask \ -H "Authorization: Bearer atlas_pat_..." \ -H "Content-Type: application/json" \ -d '{ "prompt": "What is blocking the onboarding project?" }' # Run an agent toward a goal (needs the ai:write scope). curl -X POST https://api-atlas.wrxstack.com/v1/agents/$AGENT_ID/runs \ -H "Authorization: Bearer atlas_pat_..." \ -H "Content-Type: application/json" \ -d '{ "goal": "Draft the call recap and propose follow-up tasks" }' # Approve an action the run is holding for a person. curl -X POST https://api-atlas.wrxstack.com/v1/agent-runs/$RUN_ID/approvals/$APPROVAL_ID \ -H "Authorization: Bearer atlas_pat_..." \ -H "Content-Type: application/json" \ -d '{ "approve": true, "note": "Checked the recipient" }'
Assistant vs. ChatGPT Enterprise.
ChatGPT Enterprise is a general chat assistant with connectors. Ask Atlas works inside your Atlas workspace, with scoped tools, approvals, and audit on each action.
wrxstack Assistant vs. ChatGPT Enterprise
| Capability | ChatGPT Enterprise | wrxstack Assistant |
|---|---|---|
| Where it runs | OpenAI's cloud | Atlas servers, or a model you choose |
| Model choice | OpenAI models | Atlas models, a provider with your keys, or your own server |
| Acts on tasks, CRM, contracts | Through connectors and custom GPTs | Native to Atlas modules |
| Approval before actions | Depends on the connector | Per-tool approval queue |
| Audit of AI actions | In ChatGPT's admin tools | In the Atlas audit log, next to user actions |
| Open-ended research and writing | Its core use | Focused on work in Atlas |
Common questions.
Questions security and engineering leaders ask when evaluating the Assistant.
How is "acts as the user" actually enforced?
The caller's role and module permissions apply exactly as they do in Atlas, in the app and through the API. On top of that, each agent can use only the tools it was given, and each action is written to the audit log with the person who started it.
How does memory work, and is it visible?
Memory entries have a key, a value, and a scope: the workspace, one agent, or one user. An entry can carry an expiry date. You can list entries and delete any of them, in the app or through the API.
How do approvals work?
When you set up an agent, you choose its tools and which of them need approval. When a run reaches an approval tool, it stops and waits. Someone approves or rejects it, with an optional note, and pending approvals show in one queue.
How do you keep the model from inventing facts that look like actions?
Ask Atlas retrieves context from your tasks, projects, comments, and docs and returns an answer with citations, so you can check the source. Actions go through named tools, and the ones you mark for approval wait for a person. AI can still make mistakes, so review drafts before you send them.
What limits what an agent can do?
An agent can use only the tools it was given, inside the caller's permissions, and tools marked for approval still wait for a person. You can set a monthly cost cap, pause or archive an agent at any time, and review every action in the audit log.
Can we bring our own model?
Yes. By default, Atlas private AI runs its own models on Atlas servers. You can instead choose an outside AI provider with your own keys, or point Atlas at your own model server. Keys are never stored in plain text and are not shown again after you save them.
Does the assistant work across all my apps, or just one?
Across your Atlas workspace. Ask Atlas answers from tasks, projects, comments, and docs. Agents act through the tools you give them, so one instruction can update a CRM deal and open a task. A copilot bolted onto a single app can only act inside that app. Every action runs inside the caller's permissions and is written to the audit log.
Pairs well with.
The Assistant gets sharper with every other module live. Three to start with.
Type. Watch. Audit.
A pattern-matched demo that runs in your browser on sample data. It does not call Atlas; it shows the shape of an answer and of the audit trail the product keeps.
⌘ + Enter to send · Esc to clear · Runs entirely in your browser.
Assistant response · session local
Who this is not for.
Assistant takes actions inside Atlas, so it is not a general-purpose chatbot. If you need an open-ended model for research and long-form writing, a general model provider will fit better. Atlas does not hold SOC 2 or other formal certifications. Organizations of every size, up to large enterprises, can review the security page or contact sales about the Enterprise plan with custom terms.
An AI you can actually trust.
Free to start, no credit card. Ask Atlas and agents are on paid plans, with approvals and the audit log built in.