A sales agent that lives in your Slack, works your pipeline, and cannot send anything without a human approving it first.
There is no framework here. The agent is a normal Claude Code session with the FirstTouch MCP, your HubSpot key, and a set of house rules — plus a ~500-line host that bridges it to Slack. You don't configure motions in YAML; you tell the agent what you want in a DM and it builds its own playbooks, schedules and notes inside this repo. It gets better at your business the longer it runs, because it writes down what it learns.
you: every weekday morning, work my "Q3 targets" HubSpot list — find the
right contact at each account, draft a first touch, cap it at 10/day
agent: … asks two clarifying questions, writes workspace/plays/outbound.md,
adds a schedule, stages tomorrow's first drafts for your approval
Every draft becomes an approval card in Slack — the copy, the researched reason, Approve and Deny buttons — routed to whoever the outreach sends from. Approve and the agent completes the send through FirstTouch; deny (or reply in the thread with feedback) and it cancels, adjusts, and remembers why. Nothing sends until that click. The rule is not a prompt: it's a hook that runs on every tool call the agent makes — the same control FirstTouch runs on its own internal agents — and the hook only unlocks a send for task ids a recorded human click covers.
It also knows who is talking: every message carries the sender's Slack identity, which the agent maps to their FirstTouch seat and HubSpot owner. Ask it for outreach and it drafts as you; Michael asks and it drafts as Michael, with Michael's approval card.
You need: a Slack workspace you admin · a FirstTouch workspace · a HubSpot account · a Claude subscription (Pro/Max/Team) or an Anthropic API key · Node 22+.
1. Create the Slack app.
api.slack.com/apps → Create New App → From a
manifest → paste slack-manifest.yaml → Install to
Workspace → copy the Bot Token (xoxb-…). Then Basic Information →
App-Level Tokens → generate one with connections:write → copy the App
Token (xapp-…).
The manifest matters: it turns on the Messages tab (Slack blocks DMs to an
app without it, and DMs are how you talk to the agent) and requests
files:read (so the agent can look at screenshots you send — without the
scope, Slack serves a login page instead of the image).
Slack manifests can't carry an app icon, so the agent's messages brand
themselves with the FirstTouch mark automatically (AGENT_ICON_URL overrides
it; set it empty to use the app's own avatar). For the avatar shown in the
sidebar and app profile, upload a logo once under Basic Information →
Display Information.
2. Give it a model.
On a machine where you're logged in to Claude: claude setup-token → put the
token in .env as CLAUDE_CODE_OAUTH_TOKEN. (Or set ANTHROPIC_API_KEY for
pay-as-you-go. Running locally where claude is already logged in, you can
leave both unset.) Sessions run on Opus 5 by default — the work is
judgement, and a weaker model doesn't fail loudly, it just drafts worse.
3. Connect FirstTouch.
The repo already registers the FirstTouch MCP (.mcp.json) —
nothing to add. In the repo folder, run claude, type /mcp, and authorize
FirstTouch. Claude Code owns that OAuth and refreshes it — the host never
holds a platform credential.
4. Add HubSpot (either way works):
- OAuth, same as FirstTouch: the repo registers HubSpot's MCP too — in the
same
/mcpscreen, authorizehubspotalongsidefirsttouch. One pass, no token to manage. - Or an access token: HubSpot → Settings → Integrations → Private Apps →
create one with the CRM read scopes (write scopes only if you want the agent
updating records) → put the token in
.envasHUBSPOT_ACCESS_TOKEN.
5. Run it.
cp .env.example .env # fill in the four values
npm install
npm startThe log prints a claim code. DM the bot that code in Slack — you're the operator. Then just talk to it.
For a server: on Railway, create a service from your fork of this repo (it
builds the Dockerfile), mount a volume at /data, and set the .env
values as service variables — no public networking needed, the host dials out.
The volume is the agent's disk: its memory (CLAUDE.md, workspace/,
schedules), its session transcripts and the approval records all live there,
so redeploys update the code without giving the agent amnesia.
A container has no browser for the FirstTouch /mcp authorization, so do it
once on your machine (step 3), then:
npm run seedand set the printed value as CLAUDE_CREDENTIALS_SEED on the service — the
host hydrates it on first boot. The script exports only this repo's MCP
grants: your claude.ai login and any other servers' grants deliberately stay
on your machine. The same works anywhere Docker runs:
docker build -t firsttouch-agent . && docker run -d --restart unless-stopped \
--env-file .env -v agent-data:/data firsttouch-agentInside the container the host runs as a non-root user, and the send guard is re-synced from the image on every boot: the volume holds what the agent has learned, the image holds its rules.
The agent starts knowing FirstTouch and HubSpot, but nothing about your business. Good opening moves:
- "Here's the team: … . Michael owns outbound, I own deals. Find their FirstTouch accounts and remember who's who."
- "Work my 'Q3 targets' HubSpot list every weekday at 8 — right contact per account, one thoughtful first touch each, 10/day cap, route approvals to each rep."
- "Watch our LinkedIn posts and draft follow-ups to everyone who engages."
- "Customers are companies where
planisn't empty — never touch them."
It writes what it learns into workspace/team.md and its playbooks into
workspace/, and schedules its own recurring runs in schedules.json. Ask it
"what do you know about us?" or "show me your outbound playbook" any time —
its memory is files in this repo, and all of it is yours to read and edit.
Skill packs — ready-made playbooks for common motions (founder-led
outbound, warm-engager follow-up, stalled-deal reactivation, inbound
speed-to-lead, website-visitor follow-up, customer referrals) live at
firsttouch-agent-skill-packs.
The Docker image ships all of them pre-installed; running from a checkout,
unzip a pack into .claude/skills/ and the agent picks it up.
.claude/hooks/guard-send.mjs runs before
every MCP tool call in every session, scheduled or interactive, and denies:
| Call | Why |
|---|---|
| Any immediate-send tool | delivers to a person with no queue |
| Any email send | the agent drafts; a human sends |
add_dynamic_action without isHumanApprovalRequired: true |
would send on creation |
add_dynamic_action without owner and assignee |
approving would send from the wrong person's mailbox, irreversibly |
complete_task without a recorded human Approve click |
the agent never approves its own work — the host records each click (Slack-authenticated identity) and the hook permits completing exactly those task ids |
| Publishing a flow without a recorded Approve click | live flow copy sends automatically to everyone enrolled — the agent drafts the flow, the card shows every step, a human makes it live |
Flow drafting is open (a draft sends nothing) and enrollment into
published flows is allowed; create approved-flows.txt (one flow id per
line) to restrict which ones.
The hook keys off bare tool names, so renamed or UUID-namespaced connectors
can't slip past it. It fails closed on anything it can't parse. Tests in
test/guard-send.test.mjs — treat a failure there
as a release blocker, not a flaky test.
Be honest about the trust model: inside those rules, the agent is a full
Claude Code session — it runs commands, edits this repo, and reads its own
.env. The container (or the machine you run it on) is the agent's
computer. Give it its own FirstTouch seat and a HubSpot token scoped to what
you actually want it doing, and treat the box as belonging to the agent.
- DM it like a colleague. Threads have real memory (a DM is one continuous
session; each channel thread is its own) — start a new thread for a new
topic, and say
!resetto start a surface's conversation fresh (its files and notes stay). - While it works, one status message narrates what it's doing, then becomes the answer.
- Scheduled runs post their reports to the channel you chose, with the same narration.
- Approve inside FirstTouch instead of Slack and the card settles itself — the two queues stay in step whichever side you click.
- Approval cards land in each sender's channel with Approve/Deny buttons. Clicking settles the card, wakes the agent, and it completes or cancels the send — replying in the card's thread is how you give it feedback ("shorter", "wrong persona"), and it keeps what it learns.
- In channels the agent only answers @mentions — including in card and
report threads, where people also talk to each other. DMs need no mention.
Who may talk: the operator and
ALLOWED_SLACK_USERSanywhere; anyone in the channel when @mentioning inside a card or report thread — membership in the approvals channel is the authorization. To hand the agent off, deletestate/operator.jsonand restart — a new claim code prints.
MIT.