Skeleton for Teams
What your company knows stays your company’s.
Skeleton keeps your team’s knowledge in spaces your company owns. Every AI your people use reads from it with permission, every read is logged, and you decide which models do the work.
Why this matters now
01
Models are becoming interchangeable
Every company can buy the same models. What sets yours apart is what your people know: how you price, how you sell, how you ship, and why you decided what you decided.
02
Context in the prompt leaves your control
When every record is put into the prompt, that knowledge goes to a vendor in a form nobody on your team can audit or take back.
03
Skeleton keeps it where you can see it
Your knowledge stays in your spaces. A model gets only what a task needs, every read is logged, and you can change models whenever a better one arrives.
01 · Your team
When someone leaves, their work stays in the company’s spaces.
Everyone on your team knows things nobody else does. Skeleton keeps what they write, decide and learn in the company’s spaces, so when someone moves on, the next person starts where they left off.
Tomás Reyes
Head of sales
Good at
In the company’s Skeleton
- Pricing exceptions playbookSales · 31 versions
- Call notesSales · 214 calls
Grace Adeyemi
Staff engineer
Good at
In the company’s Skeleton
- Incident runbooksEngineering · 12 runbooks
- Architecture decisionsEngineering · 38 records
Hiroshi Watanabe
Finance controller
Good at
In the company’s Skeleton
- Close checklistFinance · 64 steps
- Vendor termsFinance · 61 suppliers
Nadia Haddad
Customer support lead
Good at
In the company’s Skeleton
- Reply macrosSupport · 140 replies
- Known workaroundsSupport · 57 entries
Ben Kowalczyk
Product designer
Good at
In the company’s Skeleton
- Interview notesProduct · 52 interviews
- Component decisionsProduct · 23 records
Ruth Mensah
Chief operating officer
Good at
In the company’s Skeleton
- Decisions logLeadership · 186 decisions
- Board decksLeadership · 11 quarters
02 · On your phone
Two minutes before the meeting, the answer is in your pocket.
Ask the Skeleton app what the latest board deck said, and it reads the deck in your company's Skeleton and gives you the numbers with the slide they came from.
Ask from anywhere
The Skeleton app on your phone reads the same company spaces as your desktop.
Answers you can trust in a hurry
Every answer names the document and the slide it came from, so you can open the source before you quote it.
Only what you're allowed to see
It reads the spaces you have access to, and every read lands in the access log.
WoW sales before the 3pm
Reading Leadership in your Skeleton
I’m in a meeting in 2 min. What did the latest board deck say about week-on-week sales?
Claude reading your Skeleton
2:58Read 2 docs
From the September board deck, slide 7:
- Sales were up 6.2% week on week, the fourth week of growth in a row.
- Enterprise drove most of it: £412k of new annual revenue.
- One risk flagged: two renewals may slip into Q4.
Board deck, September
Leadership · slide 7
Tell the agent what to do
03 · Privacy and security
Company knowledge, under company control.
Skeleton hands a model only what a task needs. We test that with planted attacks and measure how much of your data leaves.
99%
fewer personal details reached the model
than putting every record in the prompt, for the same 21 tasks at two account sizes: 67 against 8,904
Higher is better
0 of 168
planted attacks reached the model
When every record was put in the prompt, all 168 did
Lower is better
Writes need approval
Agents and tools propose changes. An editor in the space approves them, or a rule your admins set.
Each tool sees only what you grant
A connection to Claude, ChatGPT or Cursor is scoped to named spaces, and can be revoked at any time.
Every read is logged
Which tool read what, for whom, from which space, sealed into a tamper-evident history.
Every change is kept
Documents and memories keep every version, with who or what made each change.
Agents run in their own cloud machines
In early access, each agent session runs in its own sandboxed cloud machine, separate from the servers that run Skeleton.
Your storage, your models
Keep your Skeleton in a database you own and choose the model for each job on your own keys.
Internal benchmark on synthetic accounts, GPT-5.6 Luna, September 2026. Pasting was simulated by putting every record in the prompt; no chat app or project was tested. No attack was detected succeeding against any approach; the count is attacks that reached the model. No outside penetration test has been done. Not yet publication ready under our own statistical checks. Not independently verified. See All Benchmarks
04 · Access
Each tool sees only what you grant it.
When someone connects Claude, ChatGPT or Cursor, they choose which spaces it can read. Nothing else is visible to it. Writes wait for approval, and a grant can be narrowed or revoked at any time.
- Revoke
Claude · Tomás
Can read: Sales, Leadership
- Revoke
Cursor · Grace
Can read: Engineering
- Revoke
ChatGPT · Nadia
Can read: Sales
- Revoke
Codex · Ben
Can read: Engineering
Access log
live
- 14:04:39
Claude Code for Weekly digest agent read Customer calls this week
Sales
- 14:04:02
Claude for Hiroshi read Vendor terms
Finance
- 14:03:25
ChatGPT for Ruth read Q3 board deck
Leadership
- 14:02:48
Cursor for Grace read Incident runbooks
Engineering
- 14:02:11
Claude for Tomás read Pricing exceptions playbook
Sales
05 · Access log
Every read is on the record.
The access log shows which tool read what, on whose behalf, from which space, and when. Entries are sealed into a tamper-evident history, so the record holds up when someone asks.
06 · Private cloud runner
Early accessEach of your agents works in its own private machine in the cloud.
The cloud runner is in early access. Each agent session runs in its own sandboxed machine. It reads only the spaces it was granted, every read lands in the access log, its network traffic goes through a filtered proxy, and what it produces comes back to your team for approval.
- ›session started · private machine · sandboxed
- ›granted spaces: Sales, Support · access logged
- ›task: this week's renewals and at-risk accounts
- ›read 42 call notes and 9 support threads
- ›outbound traffic through the filtered proxy only
- ›drafted Monday's renewal digest
- ›sent to Sales as a proposal · waiting for an editor
Model
Claude Code
Opus 5
Codex
GPT-6 Astra
Open weights
DeepSeek V4.1 Flash
NewSame context, same approvals, whichever model does the work.
07 · Shared spaces
One place for what the company knows.
Policies, playbooks, decisions, customer history and the work your agents produce live in spaces your workspace owns. People join a space as a viewer, contributor or editor, and what they add stays with the company.
Sales
Editor
- Pricing exceptions playbook
- Call notes
- Renewal calendar
Engineering
Contributor
- Incident runbooks
- Architecture decisions
- Migration checklist
Leadership
Viewer
- Board decks
- Decisions log
- Supplier contracts
08 · Agents and routines
Agents that work for the team, on the team’s terms.
Give an agent a name, a face and a contract, and let it run on a schedule. It reads the spaces it was granted, and everything it wants to change arrives as a proposal an editor approves.
Weekly customer digest
Routine · Mondays 8:00 · Sales
Update pricing policy, section 3
Agent · Tomás's request · Sales
Wren is running in your team’s cloud runner · Checking this week’s renewals
Early access09 · Your models, your storage
Pick the model for each job, and keep your data where you choose.
Pick which model runs each job, on your own keys: Claude, GPT, or open-weight models, and switch when a better one arrives. Keep your Skeleton in a database you control, and export your records whenever you want.
Your keys, per job
Each job, from search to agents, uses the model you assign and the key you provide.
Your database
Keep your Skeleton on storage you own, so your records sit where you chose to put them.
Take your records with you
Export your documents, memories and records at any time.
What reaches the model
Personal details sent to the model over the same 21 tasks at two account sizes. Skeleton sent 99% fewer than putting every record in the prompt, and got every task right.
Lower is better
42 of 42 tasks right
41 of 42 tasks right
34 of 42 tasks right
Accurate on the model you choose
Questions answered correctly with Skeleton, against one search over the same data.
Higher is better
Claude Opus 5
GPT-5.6 Luna
GPT-5.6 Sol
Claude Haiku 4.5
Gemini 3.8 Flash
Internal benchmark on synthetic accounts, personal details on GPT-5.6 Luna, accuracy across 360 graded answers per approach on Claude Haiku 4.5 and GPT-5.6 Luna, September 2026. Pasting was simulated by putting every record in the prompt; no chat app or project was tested. One search is cheaper and faster per question. Not yet publication ready under our own statistical checks. Not independently verified. See All Benchmarks
10 · Evidence
98.6%
of questions answered correctly
against 75.3% with every record put into the prompt
Higher is better
99.5%
on questions that connect records
against 51.0% with a single search over the same data
Higher is better
10.6x
cheaper per correct answer
than putting every record into the prompt
Higher is better
Internal benchmark on synthetic accounts, Claude Haiku 4.5 and GPT-5.6 Luna, September 2026. Pasting was simulated by putting every record in the prompt; no chat app or project was tested. A single search is cheaper and faster per question. No outside penetration test has been done. Not yet publication ready under our own statistical checks. Not independently verified. See All Benchmarks
Keep what your company knows.
Set up a workspace, bring in the tools your team already uses, and decide what every AI is allowed to see.