Nothing sends until you say so.
A new agent starts in draft-first mode. Watch the approval queue for a few days, tune the instructions, then let it send on its own. Three modes, and you can move between them at any time.
The agent writes a reply and parks it in the approval queue. CommSync sends nothing until a person approves, edits, or rejects it.
- A contact writes
- The agent drafts
- You approve or edit
- It sends
- Approval queue
Each parked draft pushes to owners and admins. Approve, edit then send, or reject. Two people can never approve the same draft.
- Flag for review
When a contact sounds upset, the agent sends its reply and also asks for a look. The conversation waits in Review until a teammate marks it.
- Take over
Pause an agent on one conversation and reply yourself. Any member can. Hand it back when you are done.
Eight checks before a single word.
Before an agent answers, CommSync runs an ordered gate. Any check that fails ends the turn, and most of them end it without a model call.
- Quiet hoursNo reply outside the hours you set, in the agent’s timezone.
- STOPNine stop words, in four languages. A STOP is permanent and never gets a reply.
- Suspected opt-out“Please don’t text me anymore” pauses the thread for a person.
- Never-reply phrasesYour own list of words that always go to a person.
- Automated sendersNo reply to a machine, so two bots never talk to each other.
- Turn limitsAt most 6 replies in a row, and 20 a day in one conversation.
- Security mailA code or a password reset is withheld. The agent never reads it.
- Link checkEvery link in a reply must have appeared in something the agent read.
- Safety check
Links, or Full. Full asks seven yes-or-no questions of every draft: money, advice, an unsupported fact, a leak, tone, a never-reply, and the draft as a whole.
- Your rules
Up to 20 plain-language rules. A draft that breaks one gets one rewrite, then waits for a person.
- Memory
One summary per agent and contact, with names, preferences, commitments, and open items. It survives a contact merge.
Your pages, your files, your own APIs.
An agent answers from what you give it, and it can act through the systems you already run.
Up to 20 sources a workspace. Upload a file up to 15 MB, or crawl a website up to 30 pages. The agent reads the six closest passages for each reply.
Your own HTTP endpoints, called mid-conversation. Four auth modes, HMAC-SHA256 signed, and a secret that only ever goes to its own host. A read retries once; an action never does.
An agent is an MCP client too. Connect servers from a catalog that mirrors the official MCP Registry, plus 14 curated business apps.
Short, normal, or detailed, per channel. Plain or simple formatting for email. Texts stay short by default.
- find_estimate_slots
- book_estimate_visit
- get_job_schedule
- reschedule_crew_visit
An estimates agent checks the calendar and books a visit. A scheduling agent moves a crew. Each calls the business’s own endpoint, and neither can invent a slot the calendar did not return.
Every conversation gets a score.
Agent Analysis reads each finished conversation against the goals you set, captures the fields you want, and tells you how it went. You correct it, and the corrections count.
- Answered
- Handed to a person
- Contact stopped
- Unfinished
- No reply needed
- Goals and fields
Up to 30 goals, pass or fail or scored 1 to 5, with weights and a critical flag. Up to 40 fields in 11 types, from money to a date.
- Review loop
Thumbs, corrections, a review queue, and CSV export. An overview compares the period before, and a weekly email sends six numbers.
- Alerts
Up to 20 alerts an agent, by push and email: a failed critical goal, a frustrated contact, a handoff.
With several agents on one line, the router picks the right one for every conversation, and a handoff passes what the last agent knew to the next.
Describe it, test it, then turn it on.
One text box drafts the agent: its name, role, instructions, two or three tests, and a goal or two. Nothing saves until you press Create. Then run the tests before a real contact ever writes.
- Dry run
The same prompt, tools, model, and safety check, with nothing sent.
- Saved tests
Three kinds: the next reply, a tool call, or a simulated conversation where a model plays the contact for up to 10 turns. Run 1, 3, or 5 times; failures group by cause.
- Versions
Every save is a version with a summary of the change and a Restore button. A test run records the version it ran on.
- Front desk
- Scheduling
- Billing questions
- Support
Build the whole agent from Claude or ChatGPT.
96 of the 223 MCP tools cover agents. An AI app can create one, draft it from a description, give it lines, knowledge, and tools, write tests and goals, read its analysis, and approve its drafts.
What an agent does, and what it never does.
Short answers before you build one.
- Which plan has agents?
- Pro has one agent and Power has five. Starter has none. A paused agent keeps its slot. Every new workspace gets 7 days of Power to try one.
- Who approves the drafts?
- Owners and admins. Members can use agents once an owner turns the switch on, and any member can take a conversation over at any time.
- Will it text someone who said stop?
- Never. A STOP writes a permanent opt-out. A softer request pauses the thread for a person, so no model decides an opt-out.
- What does a reply cost?
- Each turn draws from the workspace’s AI credits, and a turn costs well under a cent. There is no monthly turn cap.
Put an agent on the line.
Describe it in a sentence, run its tests, and watch its first drafts. Try every feature of Power for 7 days, with no card.