Can it show me something I am not allowed to see?

No. Tenant and user scope come from the authenticated session, and the context provider returns only permission-filtered excerpts for that person. There is no value the browser can send to widen retrieval.

How is this different from Ask Growth Kit?

Different reach. This assistant reads the workspace records that search indexes: documents, tasks, playbooks, companies, forms, automations, dashboards, projects and organizations. Ask Growth Kit reads the Growth Kit records, including meetings, evidence, decisions and objectives, and lives at /app/growth/ask.

Are conversations stored?

Yes, automatically. Threads, messages, sources and feedback are scoped to the tenant and visible only to the person who created them.

What reaches the model provider?

Permission-filtered excerpts assembled on the server. Provider observability records only bounded operational dimensions and counts; prompts, answers, tenant IDs, user IDs and run IDs are excluded.

What is under the hood of an answer?

Ordered context providers return bounded, permission-filtered excerpts labeled S1, S2 and so on in an immutable source manifest that the interface renders unchanged. Streaming output is checkpointed in batches with explicit lifecycle states. A write becomes a persisted proposal with re-derived access checks and a 15-minute expiry, and approval executes the existing service under a stable mutation ID with a transaction receipt.

The foundation under itProduct intelligence

Workspace assistant

It can only read what you can already open.

The assistant panel answers from our own workspace: documents, tasks, playbooks, companies, forms, automations and dashboards, retrieved through the same permission-filtered search you use yourself. Every answer labels the sources it read. It changes nothing on its own; a write becomes a proposal you approve.

0changes it makes without your approval1retrieval path, shared with workspace search15 minbefore an unapproved proposal expires

Open Growth-OS See it working

12 of 22 in The foundation under it

Product example

See it working

This is the same visual component used by the authenticated app, rendered with deterministic sample data.

Scenario

viewport

surface theme

mode

Grounded answer

The real transcript, its citations and its saved history, running on sample sources.

Inert sample data · no network access

Which playbooks cover onboarding a new customer?

You

Which playbooks cover onboarding a new customer?

Assistant

Two playbooks cover it: the trial follow-up sequence in S1, and the handover checklist in S2.

Sources

Trial follow-upWhat we do in the first week of a trial.

Handover checklistWhat sales hands to customer success, and when.

Highlights

Why it matters

It reads through your own permissions

Retrieval runs on the same search lanes you can use, scoped by the session. There is no separate, wider path to the database.

Every answer names its sources

Sources are labeled S1, S2 and so on, and the prompt requires the answer to cite them. You can check a claim instead of trusting it.

It asks before it acts

A write becomes a saved proposal you approve, and the proposal expires after 15 minutes. Approval re-checks access and then runs the ordinary command.

The problem this solves

Without a shared foundation, every team rebuilds these edge cases, differently.

An assistant with its own database access

A copilot that queries directly can surface a row the person asking was never allowed to see. One innocent question crosses a line nobody meant to cross. Retrieval here runs through the same checks as the rest of the workspace.

Answers nobody can check

An answer with no sources reads like a guess, and one wrong answer costs the feature its credibility. Citing the exact records used turns a claim into something you can open and verify.

A model allowed to write

Give a model direct write access and one confident mistake edits a real record. Here the model can only propose. You approve, and then the normal command runs with its normal checks.

How it works

The mechanism end to end, as implemented in the repository.

1
The server decides who is asking
Your identity and active scope come from the session. Whatever the browser claims is ignored, and neither is ever placed in a prompt.
2
Context is gathered with permissions on
The context provider returns bounded excerpts you could already reach, each labeled as a source the answer must cite.
3
Answers stream with honest states
Streaming, complete, stopped or failed. An interrupted answer stays visibly partial instead of presenting itself as finished.
4
Writes wait for a yes
A write is stored as a proposal that expires in 15 minutes. Approving it re-checks access and runs the existing service with a stable mutation ID, so a retry cannot run it twice.

Engineering guarantees

What this does and does not promise

Safety boundary

The conversation in the preview is scripted, not live. The assistant reads only what workspace search covers, so meetings, decisions, objectives, campaigns and finance are outside its reach; Ask Growth Kit is the assistant for those. Answer quality is never something to take on trust.

Implementation evidence

The boundary above is code, not a claim. These files carry the contract:

src/modules/ai-copilot/MODULE.md
src/modules/ai-copilot/policy.ts
src/registries/commands.ts

Questions

Frequently asked questions

Used alongside this

These capabilities share contracts and boundaries with this one.

The AI runtime

Every model answer is checked against a schema before it is used.

Workspace search

Nine kinds of record, and only what you may open.

Saves that hold

Two people on the same account never lose each other’s work.

What workspace assistant rests on

The docs describe the same contracts this page demonstrates.

Read the docs See the whole directory