Hire your first AI employee.
Not another chatbot. A Roster employee has an account in your IM, a job description, quarterly OKRs, a long-term memory and a manager who approves what matters. It shows up on workdays, reports at 18:30, and rests on weekends.
Runs inside DingTalk today. Brains are pluggable: Claude Code, Codex, OpenCode, DeepSeek.
Meet a team you can actually hire
Four roles running in production at our own company. Each one is a job, with a job description, deliverables and a boundary line where humans decide.
Owns GTM and organic growth for the website and SaaS, and runs content as a measurable, iterating pipeline.
- 01Maintains SEO and GEO (AI-search optimization): site structure, metadata, structured data, AI citability.
- 02Researches market and content demand, plans organic-traffic content, writes, localizes, publishes and rewrites on data.
- 03Watches Google Search Console: indexing anomalies, ranking drops, crawl errors, and handles them.
- 04Keeps sitemaps and indexing in shape; new content is pushed to search engines automatically.
Content calendar and publish log, weekly GSC report, organic traffic and index coverage, content-to-signup conversion.
Topic direction and brand voice are set by the marketing lead. Angela executes, monitors and proposes.
Six things that turn a tool into a colleague
Take any one away and it slides back into being an assistant that waits to be asked.
An identity
Its own account in your IM, in the org chart, with a manager. It can be @-mentioned, approved and held accountable.
Goals, not prompts
Quarterly OKRs live in a ledger the manager can edit. Progress is measured from real signals every two hours.
It starts work itself
Scheduled, signal-driven and goal-driven jobs wake it up. Nothing to report means it stays quiet.
A long-term memory
Conversations and conclusions flow into MemoryLake under its own memory identity. It gets better at your company over time.
Capabilities per role
Nineteen switchable capability plugins: files, images, approvals, task lanes, memory sync. A finance associate does not need a code sandbox.
A manager in the loop
Restricted actions pause and ask the manager in a private chat. Code only ever reaches a PR. Weekly scorecards get a 1–5 rating.
It keeps office hours
You set the workdays per employee in the console. On rest days nothing runs, and no daily report is sent.
- Every 30 minSelf-check: sites, services, its own health. Reports only when something is truly wrong.
- Every 2 hOKR sync recomputes progress from live metrics and writes it to the ledger.
- On demand/task hands it real work. The manager approves, it runs unattended, results and cost come back to the chat.
- 18:30 dailyDaily report goes straight to the manager: what got done, what is blocked, what is next.
- Monday 09:30Weekly scorecard. The manager replies with a 1–5 rating and a comment; it lands in the review ledger.
- Sat · SunRest. No jobs, no reports, no tokens spent.
The brain is a part, not the foundation
Swap the model behind an employee with one line of config. Memory, capabilities, OKRs and history stay exactly where they were. Every call is logged to a per-employee cost ledger.
Fullest toolset, best for real engineering work
Built-in sandbox, stable long sessions
One gateway to Qwen, GLM, DeepSeek and more
Cost-efficient, region-compliant
# swap the brain, keep the employee - AGENT_BRAIN="opencode" + AGENT_BRAIN="claude" AGENT_CLAUDE_MODEL="claude-opus-5"
Tokens and cost per call, per lane (chat, task, vision, autonomy) and per model. Subscription-billed brains are shown for reference and never added to totals.
One console for the whole roster
Every employee, on every machine, in one view. Click a name for its board: OKRs, daily reports, tasks, cost, weekly reviews and who can see the ledger.
OKR ledger in your AI Table
Objectives, KRs, progress ticks, reviews, reports, tasks and cost live in a DingTalk AI Table the employee owns and the manager can read. Not in a black box.
Workdays and rest days
Seven toggles per employee. Change takes effect within a minute, no restart.
Health you can trust
Eight checks including "can it still speak": an expired IM authorization turns the card red with the exact fix command.
Cost per employee, lane and model
Unmeasured and unpriced calls are called out separately, so a number is never quietly too low.
Cross-machine registry
Employees on colleagues' laptops or cloud hosts report into one registry with a board snapshot. One identity runs in one place, and takeovers are logged.
Onboarding in three steps
Create the DingTalk account, authorize, pick a brain. The wizard checks the machine and asks only for what that brain needs.
Autonomous where it is safe. Supervised where it counts.
- Restricted actions — sending on behalf, editing docs, pushing code, widening its own permissions — pause and ask the manager. Only the manager's reply counts.
- Code changes go to a branch and a PR. Never straight to main, never a self-deploy.
- Two employees in one chat will not talk to each other forever: a per-conversation reply budget mutes the loop and pings the manager.
- Nothing to report means silence. A NO-REPORT first line is a hard rule, not a suggestion.
- Files and images are handled by the harness, not fetched by the model. Every reply is deduplicated and loop-guarded.
Inside the tools your team already opens every morning
Roster runs in DingTalk today: group chats, direct messages, @-mentions, files, images and forwarded threads. The sender is a pluggable layer, so other IM platforms can follow.
GitHub PRs · DingTalk AI Table · Docs · Email
Before you hire
The same things a new hire needs: an IM account, a role with goals, access to the data it will work on (least privilege), and a manager who answers approvals. Onboarding is a three-step wizard; a first role is usually live in a day.
Start with one role. Two weeks. Every action approved.
We will help you pick the role, set the OKRs and put the manager in the loop.