Connect your AI agent to Odiverse

Odiverse is built agent-native: external AI assistants and automations can operate a company's back office through first-class agent APIs, governed by the same permission model, audit trail and human-approval gates that apply to every human user.

Remote MCP server

Odiverse exposes a remote Model Context Protocol server. MCP-capable assistants (Claude, custom agents, IDE integrations) connect over HTTP and receive tools (whoami, invoice queries, draft creation, readiness checks and the full read catalog), filtered by the authorizing user's permissions before the model ever sees them.

Authorization is standard OAuth 2.1 with PKCE and Dynamic Client Registration (RFC 7591): an unauthenticated call to the MCP endpoint returns 401 with a WWW-Authenticate pointer to the resource metadata, so compliant clients can register, obtain user consent and connect without manual configuration.

The security model: agents are never above humans

• Scoped agent tokens: an agent's effective permissions are the intersection of its granted scopes and the authorizing user's permissions. No scope can exceed what the human could do
• Human approval queue: sensitive writes attempted by an agent (issuing invoices, paying payroll, sending remittances) don't execute: they queue for a human, who reviews and approves. The action then runs with the human's authority
• Hard blocks: irreversible legal actions (tax filings to the authorities) are refused to agents outright, queue or no queue
• Idempotency keys on writes, per-agent rate limits, and an audit trail that records the origin of every action: human, Odi, or which external agent
• Instant revocation: agent tokens are managed (and revocable) from the app's configuration by the account's humans; agents cannot create or revoke tokens themselves

CLI

The odiverse command-line client generates typed flags from the live OpenAPI spec. The full read catalog is scriptable, and a companion odiverse-mcp binary runs the same tool surface as a local MCP server for desktop assistants.

Also available

• OpenAPI: the complete REST surface is described by a live OpenAPI spec; read operations are exposed through MCP and the CLI automatically
• Microsoft 365 / Copilot connector: generated from the same spec, authenticating against the same OAuth server
• llms.txt: a machine-readable product summary at odiverse.com/llms.txt (and llms-full.txt for the long version)

Why this exists

Inside Odiverse, Odi already runs the back office as an AI teammate, with 55 permission-filtered tools, previews and confirmation gates on every write. The agent platform extends that same architecture outward: your Claude, your automations and your custom agents get a governed, auditable way in, instead of screen-scraping or shared passwords.

The deeper architecture (GPU for interpretation, deterministic CPU code for accounting and tax) is covered in our guide to AI agents in accounting and the AI accounting agent page.