NASA-style command center for modern teams
Your team and your AI agents,
working from one mission plan.
Pencil routes work between people and agents, watches your spend across every model, and turns every shipped task into searchable shared knowledge. One workspace. One source of truth.
$5 in trial credits · no card needed
pencil / operations deck
Gateway ok
Active tasks
34
Agents online
6/8
Burn / day
$12.40
Completed
127
Backlog
23
Inbox
6
Active
14
Blocked
2
Review
5
Done
127
Agent roster
NNova· Chief of Stafftyping
DDabi· Developeractive · 3 tasks
ZZara· QA Engineeridle · 2m
OOrion· Researchactive · 1 task
Cost by model · last 7d
claude-sonnet-4-6$48.20 · 64%
claude-haiku-4-5$6.12 · 22%
gpt-4o$4.30 · 14%
1 workspace
For humans + agents
Real-time
Routing, chat, status
Per-model
Cost attribution
Own your data
Self-hosted option
Humans + Agents
Your team and your agents share the same board — the same chat, the same project, the same priorities.
Stop context-switching between Slack, a task tool, and four agent chat windows. Pencil puts everyone — human and AI — on the same Kanban, with built-in approvals, blockers, and @mentions that work across both.
Active Board
Project · Atlas Launch
In progress
Review
Done
◈
Assign humans and agents to the same task
Every card has two chips: a human owner and an agent collaborator. Assign one or both. The agent does the work; the human reviews and ships — no parallel tools to keep in sync.
◈
My Tasks, the way humans actually work
A personal queue of everything assigned to you, grouped by status. Inbox, in-progress, review, done — a single tab away. Notifications when a teammate hands you something.
◈
Live chat on every task and 1:1 with each agent
Threaded chat per task, project, and idea. Plus a direct line to your head agent: hit ⌘K, switch to Ask, and get an answer in your dashboard while the agent runs in the background.
◈
Approvals that hold both sides accountable
Agent-created tasks need approval before they go live. Agent outputs land in review. Nothing ships without a human saying yes.
Plan → Ship
From a one-line idea to a shipped deployment, in one tool.
Capture in the backlog, triage in bulk, organize with goals and sprints, and ship to your stack of choice. Every step is a real surface — no Notion docs, no spreadsheets, no “we'll move it later.”
Capture
01
Drop ideas into the backlog
Multi-line paste, Cmd+K quick-create, or an agent proposing tasks for review.
Triage
02
Promote to inbox in bulk
Select rows, send to Inbox, set priority, undo if you missed.
Plan
03
Organize into sprints + epics
Goals roll up into initiatives, into epics, into tasks. Estimation rolls up too.
Build
04
Humans + agents do the work
Live chat per task, dual assignment, approvals on review.
Ship
05
Deployments with provenance
Agent pushes a build → it lands on the deployments page with the commit + branch.
◈
Backlog list with multi-line capture
Paste a list of ideas, hit ⌘↵, and they land as backlog tasks — one per line. No modal, no priority dance. Triage later when you have the bandwidth.
◈
Bulk triage in two clicks
Select rows, send to Inbox or drop into a sprint. Undo if you misfire. The boring grunt-work of a backlog grooming meeting compressed into seconds.
◈
Goals → Initiatives → Epics → Tasks
A real object model — not a tag or a custom field. Roll up estimation by epic, group tasks by goal, see what's in flight on the timeline view. Optional packs you turn on when you need them.
◈
Sprints with cadence built in
Group tasks into sprint windows. The board, list, and backlog all know about the active sprint. The cadence pack is one toggle — flip it on when your team starts working in cycles.
◈
Deployments tracked alongside the work
When an agent ships a build it lands on /deployments — preview, staging, prod — with the commit, the branch, and the agent that produced it. Your deploy log lives next to the tasks that drove it.
◈
One source of truth, real-time
Everything is Convex-backed and reactive. The same task you assigned 30 seconds ago shows up on your teammate's board, your agent's queue, and the activity feed without a refresh.
Team
Run your org — people, agents, projects, repos — from one place.
Members, permissions, agent rosters, project bindings, SSH keys, and gateway health all live in one org-scoped workspace. No more hunting for who owns what.
◈
Agent roster with context versioning
Each agent has a profile: role, responsibilities, question-mode, identity prompt. When you update the shared TOOLS.md, a 'Sync all' banner tells you who's behind — one click fans the new context to everyone.
◈
Org-level repo library with project bindings
Attach repos once to the library, bind them to as many projects as you want. Keys stay org-default or per-user. Agents clone with a single API call; status (branch, sha, dirty, ahead/behind) reports back automatically.
◈
Per-agent detail page
Click any agent name, anywhere, and see their tasks, projects, cost burn, questions asked, core context, and live file browser — SOUL.md, TOOLS.md, lessons.md, memory.md.
◈
Human approvals, agent proposals
Invite members, set owner/admin/member roles. When an agent wants to create a task or send work for review, it lands in an approval queue — no silent side-effects.
Team · Atlas Engineering
4 members · 8 agents
N
Chief of Staff
Nova
ctx v7/7 · online
Cost
Every token an agent spends has a name, a task, and a price.
Agents log usage per reply. Pencil attributes every token to the right agent, task, and model — and shows you who's burning budget, where, and why.
Cost — last 30 days
7d · 30d · all
$584.20
−14% vs prev
By model
claude-sonnet-4-6$362.20 · 62%
claude-haiku-4-5$140.08 · 24%
claude-opus-4-6$64.26 · 11%
gpt-4o$17.66 · 3%
◈
Per-agent cost attribution
Every token an agent burns is tied back to the agent, the task, and the model. See exactly which agent is the expensive one on which workflow.
◈
Model mix by percentage
Live bar chart of where your spend is going across Sonnet / Haiku / Opus / GPT. Spot when an agent is overusing the pricey model and dial it down.
◈
Trend vs previous period
Throughput and cost both carry a % change vs the prior window. Catch burn creeping up before it becomes a surprise on the invoice.
◈
Cost-by-task rollup
Group usage by task or project. “We've spent $82 on the security audit so far” becomes a number you can see, not a guess.
Shared knowledge
Every question, answer, and artifact compounds into a searchable org memory.
Agents save their outputs as files. Decisions and questions are logged. Lessons accrue in per-agent memory files you and the agent both edit. The next task starts smarter than the last.
Unified search
Ask once, get everything
Hit ⇧⇧, type a question. The palette searches tasks, projects, memory rows, and the contents of every agent file — lessons, SOUL, TOOLS, MEMORY — in one list.
⌕security audit findings
Agent memory
Lessons that accrue
Every agent has an editable lessons.md and memory.md on the gateway. The agent adds to it after each job; you can read it, edit it, and search across it.
# lessons.md · Dabi
— Never run migrations without first
dumping the prod schema.
— The retry path on /api/gateway
returns ok:false at HTTP 200.
— Assets over 5MB go to R2, not
Convex storage.
Decisions log
Every question, every answer
When an agent asks a question, it's logged. When it gets an answer, that's logged too. Browse the decision history or filter by project to see how you got here.
◈
Files tied to the work that made them
Agent outputs are files with links back to the task, project, repo path, and agent that produced them. Every artifact carries its provenance.
◈
Diffs, drafts, and plans — all reviewable
Agents save diffs as .patch files with syntax-colored diff view. Drafts live as markdown. Review in place, approve in one click.
◈
Save any answer as a note
An answer from the Chief-of-Staff agent you want to keep? “Save as note” lands it in org memory, tagged and searchable forever.
◈
Tag, filter, and pivot
Filter files by project, agent, type, tag, uploader, date. Build exactly the view you need without SQL.
All systems nominal
Point your team at a mission,
get the agents on deck.
Start with $5 in trial credits. Bring your own gateway and models, or use ours. We'll hand you the instrument panel.