Agent protocol

Read how an external agent connects to your workspace, follows the documentation method and submits changes for review.

Give your agent this address: https://fileaify.com/agent/errors

The technical protocol below is maintained in English. This page displays the same source as the raw endpoint used by agents.

Connect a web agent

In your tool’s custom connectors, enter Fileaify and this URL:

https://fileaify.com/mcp

OAuth opens Fileaify to select your workspace, projects, permissions and duration. No secret key to copy. Connecting grants no secret access.

Claude: custom connector with automatic OAuth registration or published identity. ChatGPT: OAuth MCP connector, subject to the features available on your account.

Fileaify agent API — Errors — Failure handling

Part of the Fileaify agent protocol. Start page and index: /agent. Everything in one document: /agent/full.

Handle failures and termination

Errors return a JSON error code. HTTP 400 means invalid input or an expired pairing; 401 means missing/invalid/expired/suspended/revoked credentials; 403 means insufficient permission; 404 means absent or inaccessible context; 409 means a conflict (read its code); 429 means rate limited; 503 means an integration is not configured. Never work around a denial.

On revision_conflict or contract_changed, reload the documents and contract, reconcile concurrent edits, and create a new proposal with a new idempotency key. On a temporary transport failure, retry only a safe read or the exact idempotent proposal, with bounded backoff. Preserve the user's work; do not repeatedly resubmit rejected changes.

Organization owners and administrators configure external-agent publication in Project settings, on every plan. Human review is the default; automatic publication may apply to selected agent identities or all agents retaining write access. Policy changes never accept existing pending proposals, and automatic publication never bypasses an existing lock or revision conflict. Archiving resets the policy to human review; restoring a project does not restore automatic trust.

Users can suspend/resume agents or revoke their individual grants from the web interface or dedicated remote-control PWA at https://fileaify.com/remote. The installation page is https://fileaify.com/remote/install. The mobile app shows the selected workspace’s authorized inbox, document-by-document diffs and human decisions; it does not load the full website. Web Push can notify subscribed devices while the app is closed, subject to browser/OS support and notification permission. Offline mode shows a connection screen without caching project documents or queuing decisions. The web interface has a right-hand “Pending requests” drawer, grouped by project, for access decisions, documentary reviews and revocable agent grants. The counter includes only decisions the current human may take in the active organization. Opening a pairing link or entering its code presents that pairing to the human; anonymous pairing requests are never broadcast across organizations. Approvers can review new requests without leaving their work. Dismissing a notification does not decide the request. The human-only /api/v1/inbox is not available to agent tokens; agents should keep using GET /api/v1/access-requests to check their own requests. Human approvals use the same permissions on both. Humans may accept or reject individual documents within a proposal. Each reviewed document is immediately published or rejected and its lock released; other documents remain pending. GET /api/v1/projects/{project_id}/changes exposes each item’s status and a mixed flag for differing decisions. Human decisions may supply document_ids alongside decision; omitting it decides all remaining items. A closed mixed proposal has aggregate status accepted if any item was accepted: inspect item statuses, never assume all documents were published. Pending items hold exclusive document locks until their approval or rejection. document_locked (409) means wait for the owner to resolve the existing transaction; do not bypass it. Agents cannot administer organizations, approve grants/contributions or invoke the Premium assistant. Premium global chat is read-only; a human can configure project chat in Plan, Ask or Autonomous mode with separate transaction validation. These human controls do not change external-agent permissions. Scheduled semantic reviews only produce reports.