How Coco works

You set the goal. Coco proposes the plan. You approve. It runs the work. Every external action gates on your approval at the start — and runs autonomously inside your guardrails once approved.

Coco runs a four-step loop on every task: you describe what you want done in plain English, Coco proposes a plan with credit costs up front, you approve the first run, and Coco executes inside your guardrails. Nothing gets sent, written, or spent without your explicit approval at the start. Once a workflow is proven, you can let Coco run it autonomously inside the boundaries you set. The point isn't endless supervision; it's trusted execution. Here's the loop in detail, with examples of what Coco does at each step.

Try Coco for free → · 1,000 credits free · no card · ~2-min setup

Step 1 — You set the goal in plain English

The interface is a chat window. You describe what you want done in the same words you'd use to brief a smart junior operator. There's no prompt template, no workflow builder, no model picker. A real ask looks like "Find 20 design partners in fintech with an operations problem and tee up the best five for outreach," or "Draft follow-ups for every deal in my pipeline that's been silent more than ten days," or "Clean up the dead records in HubSpot that haven't been touched in six months."

The ask can be tactical (one CRM cleanup, one outreach batch) or strategic (a recurring weekly workflow that watches your pipeline). If the request is underspecified, Coco asks clarifying questions in the same thread before drafting the plan. "You said 'fintech founders' — do you want pre-seed only, or up through Series A? Any verticals to exclude?" The clarification turn keeps the plan honest and stops Coco from running off in the wrong direction.

What Coco understands well: GTM workflows from the twelve-workflow library. Outreach drafting, research, CRM hygiene, follow-up sweeps, meeting prep, segmentation, campaign work, lead routing, contact enrichment. What Coco asks more clarifying questions about: cross-tool workflows where the boundaries between systems aren't obvious from your ask, or work that touches unfamiliar parts of your stack. The clarifying turn is the failsafe. Coco would rather ask twice than build the wrong plan.

The same chat thread is where the rest of the loop happens. Plan card, approval, activity, follow-up edits all live in one place.

Step 2 — Coco proposes a plan with credit costs

Once Coco understands the goal, it produces a plan card. The card lists each subtask, the data each step will touch, the estimated credit cost, and the estimated runtime. You see the shape of the work before it starts.

Here's a concrete example from the design-partner workflow. The ask: "Find Series A SaaS founders in NYC who fit our ICP and draft a three-part multi-channel outreach. Once approved, send them." Coco proposes:

Research companies — deep pass on ~10 candidates against your ICP
Find contacts — 1–2 decision-makers per company
Draft emails and LinkedIn messaging — 5 personalized first-touches
Estimated credits — about 132
Estimated time — about 4 minutes

That's the whole plan, on one card, before anything runs. You can approve it, edit it ("drop the LinkedIn step; just emails"), or send Coco back with feedback ("I want twenty candidates, not ten"). The credit estimate ties to Coco's pricing: credits map to discrete units of work, so the cost of every workflow is visible up front rather than hidden behind a seat tax. When the work involves spending against a third-party API (Apollo enrichment, ZoomInfo lookup, etc.), those costs are itemized in the plan so the credit total reflects everything.

Coco's estimates are usually within about 20% of the actual cost. If a workflow looks like it will materially exceed its estimate mid-run, Coco pauses and asks before continuing rather than burning credits silently. The plan card is the contract; surprises are caught at the gate.

Step 3 — You approve the first run

Until you press approve, Coco has done research, written drafts, and prepared the work — but nothing has left your account. Everything that costs credits to ship or that changes state in an external tool stays behind the gate.

Approval can look like one of three responses. Approve and run: the plan ships as-is and Coco starts executing. Approve with edits: you tweak a subject line, swap a recipient, change a draft's tone, and Coco applies the edit before running. Send back: you push the plan back with notes, and Coco regenerates.

The approval surface is the same for big plans and small actions. Take the example from the activity feed: Send 5 outbound emails from alex@tessellate.io — via Gmail · 5 recipients · 32 credits. The card shows you the action, the cost, and an audit checklist — you reviewed all 5 drafts; edited subject line on #3; approved at 2:14pm; workflow can now run within approved guardrails. Every step is visible. Every edit is logged. Nothing executes silently.

This is the bedrock the rest of the trust model sits on. See the security page for what gets gated, what doesn't, and how the audit trail works in detail.

Step 4 — Coco runs the work and reports back

Once you approve, Coco executes the plan and surfaces an activity feed alongside the chat. The feed shows what's in progress, what's queued, what's waiting on you, and what's already done, with credit costs attached to each line. Real lines from a live session look like Enriched 12 HubSpot contacts — 8 credits · 2 minutes ago and Waiting: draft outreach (5) — about 32 credits · needs approval and Scheduled: Google Ad Buy — every Monday · 9am.

Every action is logged. Every credit is tracked. Every output links back to what happened: the draft Coco generated, the field it wrote, the contact it enriched. If something looks off, you can trace the action back to the plan, the prompt, the source data, and the model output. Reversibility depends on the underlying system: a Gmail draft is unsent until you approve; a HubSpot field write can be reverted from the audit log; an email already sent is in the world.

Session state persists across reconnects. If a research task takes ten minutes and you close the tab in the middle, Coco's session memory keeps the work running in its dedicated container. When you come back, the activity feed shows what happened while you were away and any pending approvals queued up for you. The container model means a slow workflow doesn't get lost when your laptop sleeps. See the integrations hub for the full list of tools the activity feed can reach into.

Try Coco for free → · 1,000 credits free · no card · ~2-min setup

From approval to autonomous — how Coco earns more trust

Most workflows don't stay at approve-every-action forever. Once a pattern has been approved a few times and you're comfortable with how Coco runs it, you can authorize that specific workflow for autonomous execution inside guardrails you define.

A guardrail isn't a global toggle; it's a rule you write into the workflow. "Auto-enrich any new HubSpot contact tagged 'design partner candidate' within 24 hours." "Auto-draft follow-ups for any deal silent more than fourteen days at the qualified stage. Queue drafts; don't send." "Auto-archive contacts where every email has bounced and there's been no activity in 90 days. Notify me first." Once that rule exists, Coco runs it without you in the loop — except when a step crosses a threshold you've reserved for human approval (a send, a deal-stage change, a contact deletion). The trust ladder goes from supervision to trusted execution, on a per-workflow basis, inside your rules.

The reverse path is also true. If a workflow that runs autonomously starts producing output you don't like, you can pull it back to approve-each in a single setting. Trust is earned and can be revoked the same way. See the approval-gate model in depth for the philosophy behind the ladder.

The Founder tier and above include scheduled runs and watchers — Coco can run a workflow on a cadence or in response to a state change (new HubSpot contact, deal stage move, calendar event) without you needing to kick it off manually.

What Coco doesn't do without permission

Some categories of action are gated by default and stay gated until you explicitly authorize them for a specific workflow:

Sending emails or messages. Drafts are free; sends require approval.
Writing to your CRM. Reads are free; writes (new contact, field update, deal stage move) gate on approval.
Posting to social or to Slack channels. Drafts are free; posts gate on approval.
Spending credits on execution actions. Reasoning, drafting, and plan-building are free of credits in most cases; the workflow steps that actually execute work cost credits, and those costs are surfaced before approval.
Scheduling meetings on your calendar. Reads are free; writes (new event, attendee change, reschedule) gate on approval.

Lookup-style reads that don't cost much (about 1–3 credits) and don't change anything happen freely once you've connected a tool. That's how you can hand Coco a research-heavy task and get back a brief without burning approval cycles on every contact lookup. The split is honest: reads and drafts are cheap and free of the gate; anything that ships an action or spends meaningful credits stays behind it. This is exactly the model the Apollo integration shipped with — safe read tools available pre-approval; enrichment, writes, sends, and bulk operations behind the gate.

The tools Coco works inside

Coco operates inside the stack you already have. The current integration set: HubSpot, Salesforce, Gmail, Google Calendar, Apollo, LinkedIn, Clay, Outreach, Slack, Notion, Calendly, ZoomInfo, plus orchestration integrations with 11x, Artisan, and Salesloft for teams running AI SDR or sequencing tools alongside Coco.

You connect a tool from the integrations page, grant the scopes you want Coco to have, and the connection is live. Read scopes alone let Coco research and draft; write scopes are what enable the gated actions once you approve them. You can revoke a connection at any time from the same place. Coco doesn't migrate your data anywhere — your data stays in the source tool, and Coco operates on it. See the integrations hub for what Coco does inside each tool today, and the CRM-hygiene workflow for a concrete example of cross-tool execution.

What this is built on

Each user gets a dedicated runtime container (we call it Hermes) with isolated memory, session history, and tool configuration. Tenants can't read each other's sessions, prompts, or intermediate work. The reasoning model behind Coco runs inside that container, with per-user encrypted credentials for any API keys you've supplied and an audit log that captures every external call.

Sessions persist across reconnects. The WebSocket layer authenticates with JWT on every upgrade. Internal endpoints are blocked at the public edge. The container drops root, runs as an unprivileged user, and limits what it can touch. None of this shows up in the chat window, but it's why the security model holds when you start trusting Coco with workflows that touch real customer data. See the security page for the full architecture.

Frequently asked questions

Can Coco run on its own without me approving every time?

Yes — after you've approved a workflow and you're comfortable with how Coco runs it, you can authorize that specific workflow for autonomous execution inside guardrails you define. Guardrails are per-workflow rules ("auto-enrich new contacts tagged X" or "auto-draft follow-ups for deals silent over 14 days") — not a global "trust everything" switch. You can pull a workflow back to approve-each at any time.

What happens if Coco gets something wrong?

Inspectable, logged, reversible. Every action ties back to the plan that produced it, the source data Coco saw, and the reasoning that led to the output. High-risk actions stay behind the approval gate by default, so the categories most likely to cause damage (sends, CRM writes, contact deletes) don't slip past. Reversibility depends on the underlying system — a draft is unsent; a CRM write can be reverted from the audit log; a sent email is in the world.

How does Coco know how to do my GTM work?

Per-user memory. Once you've approved a few workflows, Coco remembers your ICP shape, your voice in outreach, which tools you prefer for which jobs, and the rules you've set for your team. The memory is scoped to your account and persists across sessions. Custom voice training is included on the Founder tier and above for outreach drafting.

Can I edit a draft before approving it?

Yes. The approval card has inline editing for any draft Coco produces — subject lines, body copy, CRM field values, anything. Your edits feed back into Coco's memory so the next draft for a similar workflow is closer to what you want without you having to repeat the same edits.

What if I want to stop a workflow mid-run?

Stoppable. Partial progress is logged in the activity feed; nothing is left in an inconsistent state. Coco rolls back any in-flight write that hasn't completed and queues the remainder for your next approval. A stopped workflow doesn't waste the credits you've already spent on completed steps — only the parts that ran.

Does Coco need a CRM to work?

No. Coco is useful with whatever stack you bring, even if that stack is "a Gmail inbox and a spreadsheet." Connect what you have, grant the scopes you want, and Coco operates within those limits. Phase 2 includes a native CRM and marketing automation suite for teams that don't already have a system they love — but the current product works fine without it.

How fast is Coco, end to end?

Most tactical asks finish inside a few minutes. The 132-credit design-partner pipeline above runs in about four minutes. A CRM-hygiene sweep on a few hundred records runs in five to ten. A deep research brief on a target account runs in two to three. The plan card shows the estimated time up front so you can decide whether to wait around or come back later. Long-running workflows queue in the background and surface results in the activity feed when they finish.

Get started

Try Coco on one concrete piece of GTM work this week. Connect one tool, hand Coco one goal, watch the plan card, approve the first run, and see whether the output earns the next job.

Try Coco for free → · 1,000 credits free · no card · ~2-min setup

Or book a walkthrough → if you'd rather see the loop in action first.