Metadata on messages, threads and inboxes
Any message, thread or inbox can carry metadata: a merge-patch of keys where null deletes. Values can be strings, numbers, booleans, timestamps, lists of strings, or a decision with a value and a probability. Changes emit metadata.updated, so other agents can react.
Declare a key once, with its type and where it applies, and every write to it is validated. A key that is a timestamp stays a timestamp, whichever agent writes it.
PATCH /v1/threads/{thread_id}/metadata
{ "metadata": {
"stage": "quote_sent",
"owner_agent": "freight-desk",
"follow_up_at": "2026-10-03T09:00:00Z",
"quote_usd": 1840
} } One call for the whole context
get_thread_context returns every message in a thread, oldest first, with safe_body text and the thread's metadata: what the agent needs to decide its next step, compact enough for a model's context.
Follow-ups without a scheduler
Set follow_up_at on a thread and nofeb keeps the timer. When the time passes, a thread.due event goes to your webhooks and the event feed with the thread, its inbox and its current metadata, so the agent that picks it up knows where the work stands without another call.
Moving follow_up_at moves the timer, and clearing it cancels it. The agent needs no cron job or scheduler of its own, and the follow-up stays on the thread where people and other agents can see it.
{ "type": "thread.due", "inbox_id": "inb_...",
"data": {
"thread_id": "thr_...",
"due_at": "2026-10-03T09:00:00+00:00",
"metadata": { "stage": "quote_sent", "follow_up_at": "2026-10-03T09:00:00Z", ... }
} } Find work by its state
Query threads across inboxes and pods by their metadata (equal, not equal, ranges, in, exists, prefix), by the probability of a decision, and by the direction and age of their last message. "Every thread at stage quote_sent where we wrote last, more than four hours ago" is one call, which is what a follow-up sweep needs. A second query finds inboxes by their metadata, as a directory.
POST /v1/threads/query
{ "metadata": [ { "key": "stage", "op": "=", "value": "quote_sent" } ],
"last_message_direction": "outbound",
"last_message_before": "now-4h" } Search that understands
Search every inbox a key can see by words and by meaning at once. Full-text search covers subjects, addresses, the new text of each message, attachment names and the text inside attachments, with field syntax such as subject:invoice; semantic search finds the message that says the same thing in other words. Filter by inbox, pod, folder, thread, sender or recipient address or domain, date (including relative dates such as now-7d), attachments, keywords, authentication failures, automated mail and typed metadata. Results carry highlighted snippets and can be grouped by thread.
POST /v1/search
{ "query": "invoice overdue payment terms",
"mode": "hybrid", "target": "threads",
"filters": { "from": "acme.example", "after": "now-30d",
"not_keywords": ["$seen"], "has_attachment": true } } Decisions on every message
Every inbound message arrives already triaged: intent, urgency, needs-reply, automated, verification email, phishing, injection risk, requires-human and language, each stored as a typed decision with a value and a probability. Add your own decision schemas beside the defaults. Rules, searches and thread queries can all use them, so an agent can skip its own triage call.
When a person corrects a decision, the correction overrides it and trains the model, and per-question accuracy and calibration stats show how far to trust each one. Decisions run without mail leaving nofeb by default; a third-party model is used only if you opt in, and sees only the cleaned body and headers.
Extractors that fill metadata for you
Receipts and orders, invoices (with account numbers masked), calendar invitations and named entities are extracted into typed metadata as mail arrives. Define your own extraction schemas and run them on your own model key, with the output validated against the schema before it is stored, or run extraction again on any message on demand.
Status of everything on this page
Every capability named above, with its status as reconciled against the engine.
| Capability | What it does | Status |
|---|---|---|
| Typed metadata on messages, threads and inboxes | Merge-patch keys onto a message, a thread or an inbox, and declare a key's type so every write to it is validated. | Shipped |
| LLM-ready thread context | A whole thread, oldest first, with safe_body text and the thread's metadata, in one call. | Shipped |
| Follow-up timers | Set follow_up_at on a thread and, when that time passes, a thread.due event carrying the thread's metadata goes to your webhooks and the event feed, so agents need no scheduler of their own. | Shipped |
| Queries across threads by metadata | Find threads across inboxes and pods by typed metadata values and ranges, by decision probabilities, and by the direction and age of their last message (every thread at stage quote with no reply for four hours), plus an inbox directory searchable by metadata. | Shipped |
| Search that understands | Full-text and semantic search over subjects, addresses, the new text of each message, attachment names and attachment text, across every inbox a key can see. Filter by inbox, pod, folder, thread, sender, recipient, date, attachments, keywords, authentication failures and metadata; results carry highlighted snippets and can be grouped by thread. | Shipped |
| Decisions on every message | A default set of typed decisions on every inbound email (intent, urgency, needs-reply, automated, verification email, phishing, injection risk, requires-human and language) with a probability each, your own decision schemas beside them, human relabels that correct and retrain it, per-question accuracy and calibration stats, and confidence-gated sending for the send_when_confident tier. Decisions run without leaving nofeb by default; a third-party model sees mail only if you opt in. | Shipped |
| Built-in and custom extractors | Receipts and orders, invoices (account numbers masked), calendar invitations and named entities extracted into typed metadata as mail arrives, plus your own extraction schemas run on your own model key, with the output validated against the schema. | Shipped |