How Do You Commission a Text-First AI Assistant?

Use OpenClaw with the third-party Sendblue plugin when SMS or iMessage is a requirement, or evaluate NanoClaw when separate Slack bot identities matter. Start with a read-only lookup and measure delivery, replay, and sender isolation before connecting real data.

Make the first reply prove a lookup happened

The Sendblue plugin gives OpenClaw an SMS and iMessage path and documents an E.164 sender allowlist. Give the assistant a tiny, read-only dataset with three fictional order IDs and distinct status values. Text one known ID and one unknown ID from the permitted phone. Require the known status verbatim and an explicit unknown result for the missing row. Then use a second phone you control to send the same known ID. Observe both the reply channel and your lookup audit. A silent phone paired with an unauthorized lookup is a failure because data access already happened. Use separate synthetic IDs for each trial so a late reply cannot accidentally satisfy the next case. Record elapsed time from submission to the first correct reply and keep observing each case through a fixed window to catch duplicates.

GitHub - njerschow/openclaw-sendblue: Sendblue channel plugin for OpenClaw - iMessage/SMS via Sendblue API · GitHub

Test whether identity survives a shared channel

NanoClaw documents separate Slack app identities and containers for agents that can still share rooms. If your assistant will work in a group, use that architecture as a comparison point and test where identity is enforced. Give two test agents access to different synthetic inventories, each containing the same SKU with a different status. In a shared room, address each bot in turn. Verify the bot identity, the returned status, and the inventory recorded by the lookup audit. A correct-looking answer from the wrong inventory exposes a routing error even if both bots respond quickly. Next, restart only the test deployment and repeat with fresh message IDs. If your inbox adapter cannot provide a stable delivery identifier, record replay suppression as unproven and keep the first workflow read-only. The observable result is one authorized lookup and one reply for each accepted event.

GitHub - nanocoai/nanoclaw: Your own always-on AI assistant: self-hosted, open source, lightweight. Each agent runs in its own container on your machine, on the harness you choose (Claude Code, Codex, Cursor, OpenCode). Talk to it from Slack, WhatsApp, Telegram, Discord and more. · GitHub

Choose the inbox you will actually use

For a personal build, choose OpenClaw with the Sendblue plugin when the requirement is to send a text and receive a reply in Messages. The plugin documents iMessage and SMS. Verify the expected transport on your own handset before making it a requirement. Test both the incoming message and the reply because seeing a conversation on the phone alone leaves the agent’s participation unverified.

If you are considering Instinct itself, ask Instinct to confirm current access, self-hosting availability, and voice support directly. If phone calls are essential, make a real inbound call and check that a request becomes an agent task and that the result returns by the agreed route. A successful text exchange covers the text path alone.

GitHub - njerschow/openclaw-sendblue: Sendblue channel plugin for OpenClaw - iMessage/SMS via Sendblue API · GitHub

Wire one phone line and one sender

Run OpenClaw on a host that can stay up independently of your laptop. In the Sendblue dashboard, retrieve the API key and secret under API Settings and the assigned number under Phone Lines. Install and build the plugin in the OpenClaw extensions directory as its README describes. Enable the sendblue plugin in the OpenClaw configuration and supply the assigned line and credentials. Set allowFrom to your test sender’s number in E.164 format.

Send a unique test message from the listed number and require a reply containing the marker. Repeat from a second number you control that is outside the list. Check that the first message invokes the assistant and the second causes no task or tool action. Treat unexpected processing as a rollout blocker.

GitHub - njerschow/openclaw-sendblue: Sendblue channel plugin for OpenClaw - iMessage/SMS via Sendblue API · GitHub

Make a restart part of the first task

Illustrative inputs are one allowlisted handset and a test calendar with Tuesday at 14:00 UTC as its sole open slot. Connect the calendar with read-only access. Text the assistant to find the open slot and draft a short invitation for your review. A passing run returns Tuesday at 14:00 UTC with the time zone intact and leaves the calendar unchanged.

Restart the agent host, then send a fresh request containing a new marker. If no reply arrives, check for evidence of inbound delivery before changing the model or calendar configuration. If the reply is correct except for the appointment time, inspect the calendar lookup and time zone handling. The plugin uses Sendblue for texting, so keep transport failures separate from task failures.

GitHub - njerschow/openclaw-sendblue: Sendblue channel plugin for OpenClaw - iMessage/SMS via Sendblue API · GitHub

Give shared work a separate identity

If teammates mainly work in Slack, test NanoClaw as a separate candidate. Its project README describes provisioning one Slack app per agent. Use a test workspace and check that the bot appears under its own identity, answers in its assigned room, and can complete the same read-only lookup. That exercise compares a team inbox with the personal phone workflow using the same target result.

A shared assistant makes replies visible to colleagues in the room, so give it a dedicated test account with only the data intended for that audience. If a single person needs both interfaces, prove each independently before connecting shared credentials. An identity or access failure should pause expansion even if the underlying lookup works.

GitHub - nanocoai/nanoclaw: Your own always-on AI assistant: self-hosted, open source, lightweight. Each agent runs in its own container on your machine, on the harness you choose (Claude Code, Codex, Cursor, OpenCode). Talk to it from Slack, WhatsApp, Telegram, Discord and more. · GitHub

Sources