Keep your trade secrets out of LLM training data.

Your inbox holds your company’s crown jewels: patent drafts, unpublished roadmaps, pricing formulas, and proprietary architecture notes. Across the industry, confidential source code and internal strategy have already leaked into third-party model pipelines after employees fed work data into everyday AI tools. comail lets you automatically quarantine sensitive threads with keyword-based filters that apply no: exclusion labels on arrival—keeping your secrets safe by default until you decide a thread is okay to share.

Takes a minute to apply. No credit card required.

inbox-assistantkeyword filters → no:inbox-assistant

Quarantined on arrival · matched by keyword filter

Patent claim draft — Project Atlas inference engine

  • Filter matched keywords "patent" and "Project Atlas"
  • Auto-applied no:inbox-assistant before the agent ran
  • Withheld from search_threads — zero tokens sent to the LLM

no:inbox-assistantblocked by default

Reviewed by you in Gmail · false positive unlocked

Public roadmap blog post — final copy tweaks

  • Caught by the "roadmap" filter and held back safely
  • You checked the thread and removed no:inbox-assistant in Gmail
  • Live label check passes on the agent’s very next request

label removedsafe to share

Keyword filters automatically tag incoming threads with no:inbox-assistant whenever they mention sensitive terms. Real secrets stay out of the LLM; harmless false positives are one click in Gmail to unlock.

The problem

Once a trade secret enters a model’s pipeline, you can’t un-train it.

From semiconductor engineers leaking proprietary source code and internal meeting notes into consumer chatbots to default AI tiers that retain user inputs for model training, the precedent is clear: anything pulled into an unvetted context window leaves your control.

01

Many LLM tiers train on user inputs

Consumer and Pro subscriptions, personal harnesses, and third-party model routers frequently log prompts for human review, evaluation datasets, or future training runs by default. If an email agent runs on one of those endpoints, every thread it reads becomes provider input data.

02

Agents fetch secrets you never meant to paste

Even if you would never copy-paste a patent filing, unit-cost table, or acquisition offer into a chatbot yourself, an inbox agent running a broad search like "Q3" or "is:unread" will silently pull that thread into its prompt behind the scenes.

03

You can’t rely on tagging secrets before the agent runs

If your agent wakes up on a cron job or notification loop, an email containing proprietary benchmarks or an unannounced product codename can be ingested before you even open Gmail to see that it arrived.

How it works

Block sensitive keywords on arrival. Unlock threads when it’s safe.

  1. Turn on your agent’s exclusion label

    In deny-list mode (default: allow), no:inbox-assistant is active automatically. In allow-list mode (default: deny), turn on the agent’s blocking label from the dashboard or via the admin MCP so no:inbox-assistant overrides any sharing rule.

  2. Auto-apply no: with keyword-based filters

    Set up standard Gmail filters—or ask your comail admin agent with create_agent_filter—to watch incoming mail for words like "patent", "trade secret", "source code", "acquisition", "roadmap", or internal codenames like "Project Atlas", and apply no:inbox-assistant on delivery.

  3. Cast a wide net without disrupting your inbox

    Because an exclusion label only hides the thread from the agent—without archiving, moving, or altering it in Gmail—you can filter conservatively on broad terms like "confidential" or "pricing". Matching threads sit normally in your inbox while staying invisible to the LLM.

  4. Remove the label if a thread is okay to share

    When a broad keyword filter catches something harmless—like a public roadmap blog draft or an everyday vendor quote—just remove no:inbox-assistant in Gmail on desktop or mobile. comail checks live labels on every request, so your agent can read the thread immediately.

admin-agent — claude

> Automatically block inbox-assistant from any incoming email that mentions patents, trade secrets, Project Atlas, source code, or our roadmap.

comail-admin · create_agent_filter (agent_id: 2, action: "hide", criteria: {query: "\"trade secret\" OR patent OR \"Project Atlas\" OR \"source code\" OR roadmap"})

filter_id: "ANe1Bmj9..." · applies no:inbox-assistant on arrival

Done. New mail matching any of those keywords will get no:inbox-assistant automatically. If a thread is safe to share, just remove the label in Gmail.

With the Filters permission enabled, your admin agent creates a native Gmail filter that applies no:inbox-assistant to matching mail the moment it arrives.
inbox-assistant — personal harness

> Summarize today’s product threads.

comail · search_threads (query: "newer_than:1d")

thr_31e0: Design Team — “Icon set for Tuesday’s release”

Withheld by no:inbox-assistant: “Patent claim draft — Project Atlas”, “Public roadmap blog post”

in Gmail you remove no:inbox-assistant from “Public roadmap blog post”

> Check the public roadmap blog thread now and draft a reply approving the copy.

comail · get_thread (thread_id: "thr_31f4")

Maya: “Here is the final blog copy for the public Q3 roadmap post...”

comail · create_draft (thread_id: "thr_31f4")

draft_id: "d-4418" · saved in Gmail Drafts

The patent draft never enters the prompt. When a harmless thread gets caught by the keyword filter, removing no:inbox-assistant in Gmail unlocks it on the very next call.

Guardrails

Fail-safe across both deny-list and allow-list agents.

When turned on, an agent’s no: label always outranks co: and default: allow. A keyword filter that applies no: acts as an automatic circuit breaker before any tool response leaves comail.

Everyday triage (deny-list)

inbox-assistant

Reads general mail by default, while keyword filters automatically apply no:inbox-assistant on arrival and comail exceptions block your Trade-Secret, Patents, and R&D labels.

view deny no:inbox-assistant (keyword filters)
view deny labels: Trade-Secret, Patents, R&D
view allow default
attachments deny filenames only
send deny drafts only

Team workflow (allow-list + blocking label)

product-agent

Sees threads matching your Product label via co:product-agent, with its no:product-agent blocking label turned on so keyword-filtered threads are still held back until you review them.

view deny no:product-agent (when turned on)
view allow co:product-agent · label: Product
view deny default
attachments deny filenames only
send deny drafts only
01

no: always wins when turned on

In comail’s resolution order, no:<agent-name> (when turned on) is checked before co:<agent-name> and before the agent’s default policy. Even if a sharing rule matches a coworker’s domain, a keyword filter applying no: keeps the thread out of reach.

02

One-click release in Gmail

False positives don’t require editing rules in a dashboard. Toggle no:<agent-name> visible in Gmail, peel the label off a harmless thread on your phone or laptop, and the agent can access it right away.

03

Keep attachments out of context

Sometimes an email body is harmless, but the attached CAD schematic, financial model, or architecture deck is proprietary. Turn off attachments and the agent sees only filenames and byte sizes—never the file contents.

FAQ

Common questions.

How do keyword-based exclusion filters work?

Gmail filters match incoming mail by keywords in the subject or body, sender, recipient, or attachment presence, and apply a label on delivery. You can create these filters yourself in Gmail settings, or grant your comail admin agent the Filters permission so it can create them for you via create_agent_filter with action: "hide". Each filter applies the agent’s no:<agent-name> label automatically as new mail arrives.

What happens if a filter catches a harmless email by mistake?

Nothing in your inbox is lost or moved—the thread simply carries no:<agent-name>, which withholds it from that agent. If you glance at the thread and decide it’s safe to share, remove the no:<agent-name> label in Gmail on desktop or mobile. Because comail evaluates live Gmail labels on every tool call, the thread is available to the agent on its very next request.

Does a keyword filter also protect an allow-list agent that has a domain rule?

Yes, once you turn on that agent’s blocking label. By default, allow-list agents only use co:<agent-name>. Turning on the blocking label (from the dashboard or via set_agent_blocking_label) enables no:<agent-name> for that agent, and when turned on, no: always outranks co:—so a thread shared by a domain or label rule is still withheld if a keyword filter applied no:<agent-name> to it.

What if an ordinary thread starts discussing a trade secret three replies in?

Gmail evaluates filters on every incoming message. If reply four in an existing conversation contains a filtered keyword like "trade secret" or an internal codename, Gmail applies no:<agent-name> to the thread, and comail immediately withholds the entire conversation—including the earlier messages—from the agent. You can also apply no:<agent-name> manually at any point.

Request access

We’re onboarding beta users in small batches. Tell us what you want your agent to handle; specific use cases go to the front of the line.

Clients / harnesses (optional)

Select all that apply.

We’ll only email you about the beta. No card, no newsletter.