About the AI MCP & Connected Assistants
The Model Context Protocol (MCP) connects AI tools to workspace context, source-attributed public-record evidence, and explicit human approvals. In OrbiLattice, MCP powers both our in-app decision workflows and a bounded, read-only external connector (`POST /api/mcp/external`) so operators can query their workspace from MCP-compatible assistant clients such as Codex CLI, Claude Code, Claude Desktop, Antigravity, and Cursor.
Inputs & Provenance Boundary
Workspace-authorized records, source-attributed public facts where available, operator inputs, model settings, and identity/RBAC context form the inputs available to each tool. A tool must expose missing or unavailable evidence instead of filling the gap silently. Every external tool takes its workspace from the scoped token credential—never from a caller-supplied argument—so one token cannot read another workspace.
External Connected Assistant Capabilities (`/api/mcp/external`)
Nineteen read-only tools are served over Streamable HTTP behind scoped workspace assistant tokens. The server filters the advertised tool list to the scopes on the calling token and enforces both scope and workspace product entitlements on every call. Proposal tools return an in-app confirmation link for the operator to review and commit inside OrbiLattice without performing database writes from the assistant.
workspace.brief:read · workspace.ask:read · workspace.summary:read · workspace.activity:read
Daily Prep, Briefs & Grounded Q&A
Retrieve your daily prep brief, prioritized daily review queue, weekly workspace brief, outreach readiness check, workspace summary counts, and recent changes since a timestamp, or ask a routed question grounded in workspace records.
workspace.followups:read
Follow-Up Queue, Reasons & Review Handoffs
Inspect which leads and contacts are waiting on follow-up and why a specific record is in the queue, with a read-only reviewProposal and in-app Mission Control confirmation link that performs zero database writes.
evidence.coverage:read · workspace.sources:read
Public-Record Coverage & Source Blockers
Check verified public-record coverage ladder status and official sources for a county or place, inspect which workspace sources need attention or are blocking research, and open the in-app source review queue via a read-only deep link.
evidence.property:read · evidence.watchlist:propose
Property Evidence, Field Observations & Watchlist Proposals
Resolve a street address against verified public records with owner identity excluded, inspect recorded deeds, assessment facts, tax status, Property Intel snapshots, and field observations, or generate a read-only watchlist proposal packet with an in-app confirmation link.
How to Connect Codex, Claude, Antigravity & Cursor
Workspace owners and admins can mint a scoped token in Settings → Integrations → Connected Assistants, export it as ORBILATTICE_TOKEN, and paste the configuration snippet for their client. There is no repository to clone and no local server code to build.
Step 1 — Issue a scoped token
Open Settings → Integrations → Connected Assistants, choose the exact read scopes your assistant should see, set an expiration date, and copy the one-time token into ORBILATTICE_TOKEN.
Step 2 — Add the client snippet
Add the Streamable HTTP configuration for Codex CLI, Claude Code, Claude Desktop (via mcp-remote), Antigravity, or Cursor, then restart or reload your client.
Step 3 — Test & audit every call
Run the built-in read-only connection check in Settings or try a starter prompt below. Every tool invocation is attributed to the token and can be revoked on the next call.
Codex CLI
~/.codex/config.toml
[mcp_servers.orbilattice] url = "https://api.orbilattice.com/api/mcp/external" bearer_token_env_var = "ORBILATTICE_TOKEN"
Set ORBILATTICE_TOKEN in the shell that launches Codex. Codex reads it when it connects, so the token never goes in the file.
Claude Code
Run once in your terminal
claude mcp add --transport http orbilattice https://api.orbilattice.com/api/mcp/external \ --header "Authorization: Bearer oat_paste-your-token-here" \ --scope user
Use --scope user so the connection follows you across projects. A token pasted with a trailing space or newline is the most common cause of a failed connection.
Claude Code (file form)
.mcp.json in your project, or ~/.claude.json
{
"mcpServers": {
"orbilattice": {
"type": "http",
"url": "https://api.orbilattice.com/api/mcp/external",
"headers": {
"Authorization": "Bearer ${ORBILATTICE_TOKEN}"
}
}
}
}Claude Code expands ${ORBILATTICE_TOKEN} from your environment, so this file is safe to commit. If the variable is unset it is passed through literally and you get a 401 rather than a config error.
Claude Desktop
claude_desktop_config.json
{
"mcpServers": {
"orbilattice": {
"command": "npx",
"args": [
"mcp-remote",
"https://api.orbilattice.com/api/mcp/external",
"--header",
"Authorization:${ORBILATTICE_AUTH}"
],
"env": {
"ORBILATTICE_AUTH": "Bearer oat_paste-your-token-here"
}
}
}
}Claude Desktop cannot call a remote MCP server from its config file directly, so this routes through the mcp-remote helper, which needs Node.js installed. Restart Claude Desktop after saving.
Antigravity
~/.gemini/antigravity/mcp_config.json
{
"mcpServers": {
"orbilattice": {
"serverUrl": "https://api.orbilattice.com/api/mcp/external",
"headers": {
"Authorization": "Bearer oat_paste-your-token-here"
}
}
}
}Save in your Antigravity MCP config and reload MCP servers so Antigravity discovers the workspace tools.
Cursor
~/.cursor/mcp.json or .cursor/mcp.json
{
"mcpServers": {
"orbilattice": {
"url": "https://api.orbilattice.com/api/mcp/external",
"headers": {
"Authorization": "Bearer oat_paste-your-token-here"
}
}
}
}Use ~/.cursor/mcp.json for global user access or .cursor/mcp.json for a single project (add .cursor/mcp.json to .gitignore when a token is inline).
Starter Prompts to Try After Connecting
Daily prep & review queue
workspace.brief:read · orbilattice_get_day_prep_brief
“Use OrbiLattice to pull my day prep brief and daily review queue for this workspace.”
Follow-up queue & in-app review handoff
workspace.followups:read · orbilattice_get_lead_follow_up_queue
“Check my OrbiLattice lead follow-up queue, explain why the top contact needs attention, and give me the in-app review link.”
County coverage & source blockers
evidence.coverage:read · orbilattice_get_coverage_status
“Check OrbiLattice public-record coverage status for Camden County, NJ and list any active source blockers in my workspace.”
Propose a watchlist (read-only link)
evidence.watchlist:propose · orbilattice_propose_watchlist
“Use OrbiLattice to propose a property-criteria watchlist for Camden County, NJ and return the in-app confirmation link.”
Explicit Connector Boundaries & Unsupported Clients
Requires an existing OrbiLattice workspace and a token the operator issues in settings — this is not an open or public API
Read-only by construction: no tool creates, edits, sends, or schedules anything; proposal tools return in-app confirmation links and save nothing
Owner identity is excluded (ownerMode=exclude): no owner name, mailing address, or contact detail is returned by any tool on this surface
No automated valuation or appraisal is returned, and public-record coverage is reported per jurisdiction with no nationwide claim
Workspace product gates apply per call: Property Intel snapshots require Property Intel and field observations require Lead Engine on the workspace
ChatGPT connectors are not supported: they require OAuth, whereas this surface authenticates with a scoped workspace bearer token
Model Registry
Models are selected based on task and policy. The registry allows governed addition of providers and variants, wired to workspace AI privacy and consent settings.
Security Context
Every call carries workspace/user identity, roles, and correlationId. Approvals are enforced before sensitive actions, and every external or internal tool execution writes an auditable record.
Events & Streaming
Real-time updates stream over Server-Sent Events (SSE) and Streamable HTTP. Standardized ai.* events let the UI update feeds, show statuses, and correlate actions across surfaces.
Approvals & Audit
Approval states are persisted in PostgreSQL with actor attribution, reasons, and correlationId. Refused, scope-denied, and product-gated calls are recorded in the audit trail alongside permitted calls.
AI Insights
Outputs preserve source and provenance context, distinguish facts from operator assumptions, and report insufficient or unavailable evidence explicitly. Operators can review that boundary before approving a next step.
In-App Feature Tie-Ins
Calendar
AI prepares follow-up reminders and SLA review prompts using workflow triggers and explicit operator confirmation.
Document Center
Prepare, review, and redline documents with source-backed explanations and human approval before sharing.
Tasks
Tasks are proposed from workspace context and approvals with end-to-end correlation and audit history.
Collaboration
Mentions and activity summaries propagate via structured ai.* events to keep operators aligned.
Deal Pipeline
A separately reviewed analysis or calculator version can be attached to an exact deal after confirmation. The assistant does not create the deal or move its stage automatically.
Automation
workflow.trigger runs gated automations with human-in-the-loop approval policies and full auditing.