Brain system replication blueprint

Notion-like editable documents composed of blocks. Pages can contain text, headings, toggles, generated images, embedded databases, generated charts, files, email composer blocks, and custom components.

Airtable/Notion-like structured tables with fields, views, formulas, record pages, sensitive columns, action columns, and full cell-level history.

Google Drive-like object storage that automatically mirrors the workspace hierarchy.

Workspace
  / Pages
      / Brain demo
          / unicorn_image.png
  / Databases
      / Deals
          / Acme AI
              / pitch_deck.pdf
              / diligence_notes.md
              / founder_call_transcript.txt

A file uploaded to a page appears under that page. A file uploaded to a database record appears under that record. Permissions inherit from the linked page, record, or field.

• Create folders and pages.
• Edit rich-text blocks with Notion-style interactions.
• Generate images directly inside a page.
• Convert nearby table data into editable chart blocks.
• Rewrite or summarize selected text with installed skills.
• Insert data from any database into text or email templates.
• Prompt-edit a chart: colors, labels, chart type, title, data source.

• Create and modify databases manually or by natural language through MCP.
• Support views, filters, sorts, grouping, relations, rollups, formulas, and record pages.
• Store full cell-change logs for SLA metrics and historical formulas.
• Mark fields as sensitive for obfuscation and field-level security.
• Use action columns as a no-code trigger surface for agents.

Replaces lightweight DocSend use cases by sharing live pages with analytics.

Use Yjs, Automerge, Liveblocks, or a custom CRDT/OT layer for live page and table editing.

• Presence indicators and cursors.
• Optimistic updates with server reconciliation.
• Conflict-safe block and cell edits.
• Incremental sync instead of full document reloads.

The system needs cell-level history because SLA metrics and formulas can depend on when values changed, not just what the current value is.

CellChangeLog
- id
- workspace_id
- database_id
- record_id
- field_id
- old_value_json
- new_value_json
- changed_by_user_id
- changed_by_agent_id
- changed_at
- source: manual | import | agent | api | mcp
- agent_run_id

Every tool call should be traceable so a human or repair agent can understand what happened.

AgentRunStep
- id
- run_id
- step_index
- type: llm | tool | skill | decision | wait | human_approval
- tool_name
- input_json
- output_json
- status
- started_at
- completed_at
- error_message
- model
- token_usage_json
- credential_ids_used

1. Parse user intent from terminal, UI, or chat.
2. Search existing agents, skills, databases, and credentials.
3. Recommend reuse before creating a new agent.
4. Create missing schema or fields when safe.
5. Request or create credential authorization if required.
6. Register trigger, workflow steps, tool permissions, and logging.
7. Run a test on a small scope.
8. Persist the workflow if the user asked for repeat behavior.

1. User creates a Brain demo guest list database with name and email fields.
2. User marks name/email as sensitive.
3. User asks an MCP-connected assistant to import the event guest list.
4. Agent searches the event platform, maps attendee data, and inserts records.
5. Optional: user adds an Action column for enrichment.

1. Cell changes to Enrich.
2. Automation engine emits cell.changed.
3. LinkedIn enrichment agent loads row data.
4. Credential vault provides approved enrichment credential.
5. Agent writes LinkedIn URL, title, company, bio, and status.
6. Agent sets Action to Enriched and logs the run.

1. Recorded call transcript is stored in Call Transcripts.
2. Agent extracts company names, people names, and action items.
3. Agent searches historical emails for context: how we met, who introduced us, last interaction.
4. Agent enriches people with LinkedIn and event-platform data.
5. Agent updates People, Companies, Relationship Context, and Tasks databases.

1. Email agent detects a founder update from a portfolio company.
2. Update is stored in Portfolio Updates.
3. Ask extractor creates Requests for help asks.
4. Classifier assigns request type and owner.
5. Resolver searches CRM for relevant helpers.
6. Email agent drafts an introduction email in the right team member's inbox.

1. Task extractor reads emails and internal conversation transcripts.
2. Likely action items become Task records.
3. Telegram bot sends a task card with buttons: dismiss, reassign, complete.
4. Dismissed tasks are treated as negative examples.
5. Weekly learning job updates rules to avoid repeated false positives.

1. User writes an email inside a page.
2. User inserts variables from a database.
3. User selects sender account and recipient database.
4. System sends individual emails, not CC, with throttling.
5. Send logs are written back to the workspace.

• Who or what triggered the action.
• Which records, pages, files, and credentials were accessed.
• Which model and prompt version were used.
• What the agent changed.
• Whether a human approved the action.
• Errors, retries, cost, and token usage.

• Virtualized table rows and columns.
• Server-side pagination and cursor queries.
• Precomputed view indexes for common filters and sorts.
• Incremental block loading for long pages.
• Optimistic UI updates with WebSocket patch sync.
• Separate search index for full-text and semantic retrieval.
• Background formula recalculation and materialized fields.

{
  "name": "Guest List LinkedIn Enrichment",
  "trigger": {
    "type": "database.cell_changed",
    "database": "Brain demo guest list",
    "field": "Action",
    "new_value": "Enrich"
  },
  "permissions": {
    "read_fields": ["Name", "Email", "Company"],
    "write_fields": ["LinkedIn URL", "Role", "Company", "Bio", "Action"],
    "credentials": ["linkedin_enrichment_provider"]
  },
  "steps": [
    {"type": "load_record"},
    {"type": "run_skill", "skill": "linkedin_enrich_person"},
    {"type": "update_record"},
    {"type": "set_field", "field": "Action", "value": "Enriched"},
    {"type": "write_audit_log"}
  ]
}

if(
  hours_since(last_changed("Stage")) > prop("SLA Hours"),
  "🟢 SLA exceeded",
  "⚪ Within SLA"
)

The exact syntax can vary, but the formula engine needs functions that reference historical cell changes.