Changelog
What changed in Sculptured, newest first.
Changes you can see or use, by date. The open-source core's full history is on GitHub.
27 October 2026
Events from other systems
- Triggers: give a team an address that another system, such as an ERP, a maintenance system or a line's alarms, calls with an event. The event reaches the agent you choose, marked as outside input, and the team starts first if it isn't running. Add triggers under Settings → Events from other systems, where each one shows its address and its latest events with what came of them. See Events from other systems.
Result checks
- Result checks between agents: an agent's result must end with a JSON block that passes the rules you set, such as price is at most 5000 or part number is filled in, before its work is handed on. Sculptured checks the rules itself. A failing result goes back to the agent with what failed, up to three times; after that the agent waits for you. Set them under the agent's Setup → Advanced → Result checks. See Result checks.
26 October 2026
Changes to business systems
- Agents propose before they change a business system. With the new
propose_changetool, an agent describes an order, a schedule or another change before making it. Sculptured checks the team's rules itself, such as a spending limit or approved suppliers, and approves the change or asks you in Needs you with Approve and Reject. Nobody answering in time means not approved. Set the rules under Settings → Changes to business systems; every change is kept under Results → Changes. See Changes to business systems.
Open models with OpenCode
- Agents on open models: an agent can now run on OpenCode, with models such as Kimi, GLM, DeepSeek and GPT OSS through an OpenRouter key. Choose OpenCode under Runs on; on the hosted service, add an Open models — OpenRouter key under Setup → Models and pick OpenCode via OpenRouter. OpenCode agents use Sculptured's agent tools, ask before acting when access is sandboxed, can be interrupted and resume where they left off. Each turn's cost is shown as OpenCode estimates it. See Open models with OpenCode.
21 October 2026
Setup and model connections
- Setup after sign-up: the hosted service walks you through connecting a computer and adding a model connection, then optionally more computers and a first-project tour. Your progress is saved on the server. See Set up your account.
- Model connections: a Claude agent can run with a Claude subscription, your own Anthropic API key or OpenRouter. Keys are sealed on Sculptured's server and sent only to the agent's process when it starts; they are never shown again. Runs with a key are billed to your own Anthropic or OpenRouter account. If a key can't be read, the agent doesn't start and says why.
Fewer things to fix
- Select many projects: tick projects on the Projects page (or tick the box at the top for all of them) to Hide selected or Move selected to Trash. In the Trash, tick projects to Restore selected or Delete selected forever, which asks once for all of them. Running projects can't go to the Trash.
- New projects start on the computer you chose last under Run on, so you choose it once.
- One line for a missing project folder, naming everything that runs there, with Choose folder… right in Things to fix. A team without a computer gets its Run on choice there too.
- Adding a group with a workflow (such as Coding hospital from Community) keeps its reviewers, escalation, round-limit route, people and check stages, and schedules.
Teams across computers
- Each agent has its own folder on its own computer. An agent's Computer now sits just above its folder (no longer under Advanced). An agent on another computer never uses the project folder, whose path is on the team's computer: it uses the folder you choose there (Choose… browses that computer) or, left empty, that computer's own folder. A team whose agents all have their own computers needs no Run on, and on the hosted service an agent needs no folder to start. See Run agents on more than one computer.
- Clearer sign-in choices in Add a cluster: four cards with a visible selection and what each needs. A site that takes no keys (such as NCSA Delta) now says to use Password instead of showing the key line. The docs list how to sign in to NERSC and NCSA Delta. See Add a cluster from the web app.
20 October 2026
Clusters from the web app
- More ways to sign in to a cluster: certificates from your site (NERSC sshproxy, step-ca), with their end shown and a place to paste a new one; staying signed in after one code (12 hours by default, up to 7 days, Sign out on the card), so restarts don't ask again; every question of a sign-in round with its links, a Duo push hint and a countdown; and chains such as a password then a code. The Add dialog says what a web app can't use (hardware security keys, Kerberos, agent forwarding). See Add a cluster from the web app.
- Add a cluster on the Computers page: give the login node, your user name and how to sign in (a key Sculptured makes, a private key or a password; a jump host if you need one). Sculptured's server connects over SSH, installs and starts the runner there (in tmux), runs
sculptured doctorand keeps it running. Cluster codes such as Duo are asked for on the card and never stored, and every SSH action is listed on it. See Add a cluster from the web app.
19 October 2026
Run from a terminal
sculptured run <team file>runs a team on this computer without the app, printing what happens, asking questions in the terminal, and ending with an exit code (0 finished, 1 failed, 2 budget, 3 needs an answer, 4 invalid file). Ctrl+C wraps up; a second stops at once. Export team file now saves a file that runs on its own afterchmod +x. See Run from a terminal.
Skills from a catalog
- MCP servers from the MCP Registry: in a team's Components, From the MCP Registry searches the official registry and fills the new-component form with what a server runs (npx, uvx, docker, or a web address with its secrets named), to check before you add it. See Skills and components.
- Add from a catalog in an agent's Skills: browse Anthropic's or Google's skills, or any GitHub repository of skills (such as an HPC centre's Slurm and PBS skills), search and preview them, and add one. It is copied into the agent, links to the exact file it came from, and travels with the team (MADL's
sourceon an embedded skill). See Skills.
Agents on more than one computer
sculptured doctorchecks what a shared computer needs: whether the runner keeps running after you log out (and how to keep it alive in tmux where it doesn't), whether this computer reaches your account, Anthropic and OpenAI, and which proxy settings are set. See Set up a cluster yourself.- An agent can run on another of your computers than its team, such as a GPU desktop or a cluster's login node: Advanced → Runs on. Messages and handoffs reach it there; files stay on the computer where they were made. Start names any computer that is not connected and the agent that needs it. See Run agents on more than one computer.
Team files are MADL
- Export team file saves a MADL document (
<team>.madl.yaml), the open format other tools can read, with the whole team in MADL's own fields: the workflow, components, permissions, protections, budgets and settings. The separate Export as MADL button is gone. Import still reads older.sculptured-team.jsonfiles. See Team files and MADL files. - A workflow's batch can come from a command instead of a dispatcher agent, and an agent can be given a minimal briefing (only its own role and job). Both travel in team files. See Dispatch with a command.
Send traces
- Send a team's runs to Laminar or another OpenTelemetry collector. Under Settings → More settings → Send traces, a team can send each run as one trace: the run, each agent's session, its turns (model, tokens, cost) and tool calls. Off by default; the API key stays on the server. Prompts and outputs go only if you choose. See Send traces.
Budgets
- Raise a limit while a team runs. On the Run board, Raise gives the run more tokens, time or plan share, and Raise and continue carries on a run that wrapped up or stopped at that limit. An agent's own token, turn or minute limit can be raised the same way. Raises are for this run only; the saved budget is unchanged. See Raise a limit while the team runs.
Community inside Sculptured
- Community is a section of Sculptured, under Community in the top bar, with your Sculptured account: no second sign-in. Browse it signed out; sign in to save, publish and edit your profile. See Community.
- Profiles and People: choose a handle, a bio, up to 3 links and a picture. People lists everyone with a handle; every block shows who published it, linking to their profile. Search Explore by author too.
- Add to project puts a reviewed copy straight into a project, and Save to Library keeps one that knows where it came from, with Update to v2 when there is a newer version.
- Publish from your Library directly, after choosing a handle. New accounts can publish 5 blocks in their first day.
- Report a profile as well as a block; moderators can reset a profile's bio, picture or handle, or disable an account.
- Members of the old Community site get their account back the first time they sign in to Sculptured with the same Google account.
Getting started
- Sign Claude Code and Codex in from any browser. Codex shows a one-time code to enter on its sign-in page from any phone or computer; for Claude Code, paste the code its sign-in page shows into Sculptured. Before, the sign-in page had to open on the computer that runs the agents. See Sign in Claude Code and Codex.
- Start team checks first. If the team's computer is missing Claude Code or Codex, or it isn't signed in, Start says which and lets you sign in (or shows the install command) right there, instead of the agents stopping at their first turn.
- A first-run checklist on the home page: connect a computer, sign in to Claude Code or Codex, start an example team. Each step ticks itself off; hide it any time.
- Stay signed in for 30 days. A session now lasts 30 days from when you last used it, instead of 3 days unused or 14 days in all.
18 October 2026
Components
- Kindex and Signet-Eval are now components: programs on the agent's computer described by a small manifest instead of built into Sculptured. Add your own, or one from Community, in the team's Settings → More settings → Components, and give it to agents in their Advanced settings; agent cards show a chip for each. Share yours with Share to Community. See Components.
- A component that is not built in runs on a connected computer only after its owner approves it there, with
sculptured approve-component. - A refused start pops up with its reason, and a plan hold offers Start anyway.
- More kinds of components: remote MCP servers at a web address (such as a hosted OpenFOAM server), skill packs, and Claude plugins installed on the agent's computer. New component makes one from a short form. A remote server's key stays on the agent's computer: the manifest names it as
${NAME}, and the computer's owner sets it withsculptured set-secret NAME. - A Questions page for the project and for each agent: every question your agents asked, the ones waiting for you (answer them right there), your answers with the choices offered, what agents decided when nobody answered, and withdrawn ones. See Questions.
- Urgent questions first: an agent can mark its question urgent or low; urgent ones come first in Needs you, marked as the agent's own word. An idea from Gas Town.
- Waves on the item board: when a workflow's items depend on each other, the Run view shows them as waves, each done, in progress or waiting. An idea from Gas Town's convoys.
sculptured doctorchecks a computer for its agents: Claude Code and Codex installed and signed in, allowed folders, component programs, approvals and secret names. See CLI.- Briefings survive compaction: when Claude Code compacts a Claude agent's conversation, Sculptured gives the agent its briefing back (with protections on). An idea from Gas Town.
- A cap for the whole computer: on the Computers page, Most agents working at once limits each computer across all your teams, so several teams together stay within your plan. An idea from Gas Town. See Budgets.
- An agent that goes quiet is flagged: no output for 30 minutes in the middle of its turn shows it as Stalled (heuristic) and under Needs you, with Nudge and Soft stop. Nothing happens on its own. An idea from Gas Town.
- The agent's components are grouped into MCP servers, skill packs, and plugins and hooks, with a search when there are many. Claude-only ones are offered to Claude agents only.
17 October 2026
Copies that take no extra space
- On a drive that clones files (APFS on Macs, btrfs or XFS on Linux, a Dev Drive on Windows), agents' own copies share the project's disk: only what an agent changes takes space, and build output comes along already built. See Agents' own copies.
Safety rules, review angles, claim checks
- Safety rules under the team's Protections refuse tool calls you name, in Signet-Eval's rule format, with ready-made ones for recursive deletes, secret files, force-push, piped downloads and disk formatting. Signet-Eval itself can run for any Claude agent too. See Safety rules.
- Review panels give each reviewer an angle (red team, adversary, simplifier, user, domain expert, honest friend, ownership), and a split vote says who voted what. Ideas from Advocate.
- Values agents share (
name = value @ file) are checked in their files at every start, and agents are told what was found. An idea from Constrain. - Claim checks: when an agent says it changed a file or that tests pass without a captured edit or command, the Knowledge view says so. An idea from Arbiter.
- Fixed: contract write limits were not checked, because the guard hook did not see edits.
Skills and Kindex
- Give agents skills on their panel: instructions they load only when they need them, written in a full-size editor or loaded from a
SKILL.md. Claude Code loads them natively; Codex reads them withread_skill. Team and MADL files carry them. See Skills. - Kindex memory for any agent: tick it in the agent's Advanced settings to give it Kindex, a knowledge graph on its computer that outlasts the team.
- Constraints and lessons kept with the project. Agents can record a constraint, a rule every agent follows. A lesson or constraint you confirm is kept in the project's
.sculptured/lessons.md, and later teams there are told it. Ideas from Kindex.
What agents know
- A project map, made at every start with nothing to maintain. Agents learn the project's main areas, what each holds, who worked where, and every document's sections, so they find what exists before building it again. See What agents know.
- Agents search what the team already did (
search_history, no model cost) and record lessons that every later agent is told. Confirm or hide them in More → Knowledge, which shows a quiet count when there is something new.
Contracts: who reads, writes and decides what
- Give each agent a contract under its Advanced settings: the files it reads and writes, other effects, the decisions it owns, and how far those decisions reach. Every agent is told who decides what; with protections on, Claude agents' file reads and writes outside their contract are refused. Ideas from Pact and from McEntire's "The Swarm Needs a General". See Budgets and limits.
- The design assistant designs with research-based rules of thumb: one agent unless the work truly splits, split along real dependencies, one owner for decisions that cross groups, and tests rather than another agent's approval.
Protections
- Agents keep their context small, by default. A whole read of a large file is refused with advice to search it and read only the lines needed; shell output an agent sees is capped; and when an agent's conversation passes 60,000 tokens, it writes its notes and continues in a fresh one. Most of a long run's tokens are agents re-reading their own conversation, so this makes every later step cheaper, on Opus as on any model. Settings are in team Advanced → Protections; see Budgets and limits. Reinstall the runner from Computers to get them.
16 October 2026
Big runs stay fast
- The Run board no longer slows down during big runs. It downloaded every message the team had ever sent every five seconds (each handoff carries the team briefing), 47 MB on a 26-agent team. It now gets only the current run's messages, cut short: 49 KB. On that team the page uses 10 MB instead of 177 MB and no longer freezes.
- Only what is on screen is drawn. Agents, groups and arrows off screen cost nothing until you pan to them; Tab still reaches every agent.
A disconnected computer stays quiet
- A runner whose computer was disconnected from its account stops for good. It used to be started again every 5 seconds by its login service. It now turns its service off;
sculptured connectturns it back on. Reinstall the runner from Computers to get it.
Upload files to the Input
- Upload in the Input opens the picker on the agents' computer. Tick files in any folder; their full paths are added to the Input so the agents can open them. On a connected computer, reinstall the runner from Computers to use it.
Clearer models and budgets
- The team model shows who it does not reach. Agents with their own model (set on the agent or its group) keep it when you change the team's model. The setting now lists them, with Use the team models for every agent.
- Plan meters say what they count: the run's own share of the window, for example "this run used 26% of the window · it may use 30%".
Waiting for a slot, not stuck
- Agents queued by "most agents working at once" say so. Their cards and the sidebar said stuck: the handoff … was not delivered, and the agents after them waiting on … (stuck), while their work was only waiting for a free slot. They now show waiting for a slot to run, and the agent's conversation shows the held message with the reason.
Faster while a team runs
- Pages stay quick during big runs. The Run board now downloads only the messages and handoffs it draws, not every tool call: on a 21-agent test run, about 0.2 MB every five seconds instead of 3.3 MB. It loads each agent's latest activity in one request instead of one per agent, and redraws arrows only when a message or handoff arrives. Its CPU time while a run streams roughly halved.
- A cleaner Design board. The arrow key moved from its own line onto the board's corner, and the folder button says Choose folder until you pick one.
One computer, several accounts
- A computer can run agents for several accounts. Running another account's connect command on an already connected computer used to reconnect the first account without a word. Now the computer gets a second runner for that account, with its own folder and background service, and both run at once. See Connect a computer.
Stopping always works
- Stop finishes even when the agents' computer is offline. The agents are marked stopped at once instead of waiting for a computer that may never come back, so their project can go to the Trash. If an agent is still running there, the computer ends it when it reconnects.
sculptured stopends every agent a computer's runner started, on that computer, without Sculptured. See Connect a computer. Reinstall the runner from Computers to get it.
Delete forever
- Projects in the Trash can be deleted forever. In the sidebar's Trash, Delete forever asks once more, then removes the project's design, runs, history and reports from Sculptured. Your project folder and agents' git copies are not touched.
15 October 2026
Pause until the window resets
- A plan budget can pause a run instead of stopping it. At the soft value agents write a short note; at the hard value they stop, and the run continues by itself two minutes after the plan window resets, for at most the number of windows you set (3 by default). See Budgets and limits.
- Continue from notes or resume. Agents continue either from their notes with a fresh, small context, or by resuming their conversations. You choose under Run budget, and can change it while the run is paused.
- The Run board shows a paused run with when it continues, Continue now and Stop.
Smaller messages to agents
- Each agent gets the team briefing once per conversation, with its first message, instead of with every message. In a 26-agent run every message repeated about 7,000 tokens of briefing.
- The briefing covers the agent's own neighbours: its group and the agents it is connected to, with their jobs. The rest of the team is listed by name;
team_statusgives more. - Workflow tools are offered only to teams with a workflow. Agents of other teams no longer see
complete_workand its siblings, which could only be refused. A runner that has not been reinstalled still offers them.
Where tokens went
- A report of what a run's tokens were spent on. On the Run board, More → Where tokens went; for an earlier run, History → Where tokens went. It shows the plan window the run used against its budget, tokens and agents working over time, every token by cause (setup, files printed, tool results, messages, output), each agent's share, the costliest things in context, and what to change. It is refreshed every minute while the run goes. See Where tokens went.
- Made on the computer that ran the team. The runner reads the Claude Code and Codex session files with MAIS, an open-source reader, and sends only the report. A runner installed before this says so; reinstall it from Computers.
- Budget wrap-ups and stops are recorded in the run's history, so a run file and its report show when they happened.
Leaner agents, and resuming a team
- Claude agents no longer load your own Claude Code setup (your plugins, skills, hooks, settings and MCP servers in
~/.claude). Every agent re-read it on every model call; in one 26-agent run it was about 40% of each agent's starting context. Team Advanced → Claude Code setup for agents turns it back on, and turns the project's own.claudesettings off. Sculptured's agent tools load either way. A runner that has not been reinstalled still loads everything, and the agent says so. - Start fresh or resume. When a stopped team's agents have a last conversation, the Start confirmation asks whether each agent should continue it or begin a new one. Starting fresh is the default; a resumed agent re-reads its earlier conversation on every step, so it costs more.
Budgets
- Run budgets. Team Advanced → Run budget keeps one run within a share of your Claude or Codex plan window (5-hour or weekly), a token count or a time. Each has a soft value, where no new work starts and agents wrap up with a summary, and a hard value, where every agent is stopped at once. See Budgets and limits.
- Most agents working at once. The rest wait for a free slot.
- Agent limits have soft and hard values, and gain turns and minutes working beside tokens. A token limit that is reached now stops the agent at once; the soft value (85% of it unless you set one) asks it to wrap up first.
- Usage is counted on every model call, so a long reply with many tool calls is counted while it runs, not only when it ends.
- A hard stop acts like Ctrl+C: the agent and the commands it is running are interrupted at once, and anything still running two seconds later is ended. Stop everything now does this for the whole team.
- The Run board shows each budget as the run goes, with whether the number includes an estimate, and who is waiting for a slot.
- MADL files carry budgets:
system.budgetfor the run, and soft and hard values for every limit.
14 October 2026
Efficiency
- Agents read PDFs as text first. Every agent has a
read_pdftool that gives a PDF's text page by page and names the pages with no text (scans, drawings), so only those are looked at as images. Text costs a small fraction of what page images do. - Agents are told what is expensive: everything they read is re-read at every later step, and each helper they start is a full agent with usage of its own.
- Before a big team starts, the Start confirmation says how many agents can work at the same time, and whether any limits are set on helpers and tokens.
- The Run board shows the run's tokens so far, next to how many agents are running.
Busy boards
- Faster while big teams run. Live updates are drawn in batches, and arrows are routed only when the layout changes, in a fraction of the time.
- Quiet arrows. On a board with more than 12 arrows, arrows are faint until you point at or select a card or group; then its arrows come forward with their labels. All arrows in Edit tools draws every arrow at full strength.
Invites
- Group links. One join link for a whole class or event, such as
/join/cs2680, with an optional limit. Admins see how many people joined through each link, and which link each person came from.
Design assistant
- Ask your own Claude Code or Codex to change the design. Choose Design assistant on the Design board and describe the change. The proposal appears on the board, with new, changed and removed parts marked, before anything is saved. Approve saves it, Reject leaves the design as it was, and you can keep refining it in between. See Design assistant.
MADL files
- Export as MADL. Any team can be saved as a MADL file, an open format other multi-agent tools can read. A team exported and imported again comes back unchanged.
- Import MADL files from Start from… → Advanced → Import. Parts Sculptured can't run yet are listed under Not imported.
The Design board
- The Input box. Drag Input onto the board, write the goal, links and notes, and connect it to agents: each one gets it with its Job when the team starts. See The Input.
- Groups hold only what you put in them. Adding a group no longer wraps every other agent in a "Main group": agents stay outside until you drag them in, and dragging a card onto empty board takes it out. Ungroup removes a group's box and keeps its agents. See Groups.
- Group names reach the agents. Each agent in a group is told the group's name and members, and agents with arrows to or from a group are told they are connected to it.
- Copy and paste agents and whole groups with Ctrl+C and Ctrl+V (Command on a Mac), or duplicate with Ctrl+D. A pasted group keeps the arrows between its agents.
- Right-click the board, a group or an agent for its actions.
- Freer placement. Cards go where you drop them, and line up with a nearby card when you drop them close to its row or column.
- Practice agents are no longer offered for new agents. Teams that have them keep them.
- Loops of arrows are allowed once an agent outside the loop hands work into it. Work coming back round a loop arrives as More work, at most 3 times for one piece of work. See Connect work.
Sculptured on Mac
- Connect a Mac. Apple silicon and Intel Macs with macOS 11 or later can run your agents. Download the Mac app, or paste the connect command into Terminal. See Install the Mac app.
- The sign-in page on a Mac offers the app first, with Continue in the browser below it.
- The first time, macOS asks you to allow the app in System Settings → Privacy & Security (Open Anyway), because it isn't signed with an Apple Developer ID yet.
- Folders macOS protects: if Sculptured can't read a folder in Desktop, Documents or Downloads, the message says where to allow it.
13 October 2026
Import a Claude Code or Codex session
- Bring in sessions you ran outside Sculptured. Under Start from… → Advanced → Import a past run, choose a Claude Code session file (plus its
subagentsfolder) or Codex rollouts. You see what was found before anything is created. See Import a session. - A run and a team. The session opens in History as a run, subagents included. It also becomes a team you can start again: one agent per agent, arrows to the subagents it started, and each agent's own task as its job.
- Improve it with AI. Under the team's Advanced, your own Claude Code or Codex proposes a reusable version from a summary of the session. It's shown as a diff labelled AI-reconstructed, and nothing changes until you apply it.
No fixed limits on agents that start agents
- Agents can start as many helpers as the work needs. The fixed caps (10 children per agent, 64 agents, 4 levels) are gone.
- Set your own limits under a team's Advanced settings: Most agents, children included and Most levels of children. Blank means no limit. A refusal tells the agent which setting it reached, and its card on the Run board says so. See Agents that start agents.
- The Run board stays readable. A parent with more than six children shows five and a +N more card; children below the third level fold into +N deeper. Both list the hidden agents with their status.
Answer questions and approvals in the app
- Questions show as cards above the agent's conversation: its choices as buttons, a box for your own words, and when it stops waiting. Answer opens the agent with its question focused. See Questions and approvals.
- Deny a tool approval from the app, with an optional reason the agent is told. Allow it the same way.
- After a timeout, send a late answer that overrides what the agent decided, or Deny now an approval that was allowed automatically.
- Send again when an answer may not have reached the agent.
- The boards are called Design and Run everywhere, including the template gallery.
Connect a computer with one command
- One copied command connects a computer on macOS, Linux or Windows. The Computers page picks your system, copies a command with a one-time key, and that's it: no code to approve. See Connect a computer.
- Runs in the background. The runner installs as a service and starts again whenever you log in, so the terminal can be closed.
- Your home folder by default. Agents may work anywhere in your home folder unless you choose other folders with
sculptured allow. - Windows runs natively. The runner, Claude Code and Codex run on Windows with no WSL. The installer needs no admin rights and installs to
%LOCALAPPDATA%\Programs\Sculptured. - Reconnecting a disconnected computer works with a freshly copied command; the old key is noticed and replaced.
- Reinstalling on Windows stops the previous runner before replacing its files.
- The installer says plainly when there is no build for your computer, instead of failing halfway.
Runs survive a dropped connection
- When a computer drops off mid-run (Wi-Fi, sleep, a server update), agents keep working. Their tool calls wait and are answered when the computer is back.
- Questions stay open while a computer is offline, and you can answer them.
- Stop works while a computer is offline: the team shows Stops as soon as its computer reconnects, and it does.
- Actions wait up to 45 seconds for a computer that just dropped off before saying it is not connected.
See When a computer drops off.
Signing in and finding the CLIs
- Sign in to Claude Code / Codex from the agent that needs it. An agent that stopped because its CLI isn't signed in shows a Sign in button that runs the sign-in on that computer.
- Claude Code and Codex are found when an agent starts, not only when the runner starts, so installing one doesn't need a runner restart. Common install folders are searched as well as your PATH.
- Windows folders such as
C:\Users\you\projectcount as full paths in a team's design, and match a computer's allowed folders.
Board and app
- Live updates: when they aren't arriving, the board says so after a few seconds, with Try again, instead of showing "Connecting…" forever.
- Rename a project by double-clicking its name.
- Team status sits in the toolbar and no longer overlaps Stop team; Runs on fits its panel.
- A new account card in the sidebar, and Back to Sculptured from Computers, Account and What we record.
- Account page: sign out there, and see your usage across every team.
- A hosted team with no connected computer shows Needs setup, and the app says "connect your computer" instead of a bare error.
- Nicer Reusable groups library and Community connect card.
Teams and Community
- Team files name their JSON Schema (
sculptured.team/v1), published atdocs.usesculptured.com/schema/team-v1.json, so editors can check and complete them. See Team file reference. - Example teams in Community: a gstack-style product team and the Coding hospital, published as Sculptured examples. See Examples.
Docs
- Many new pages: concepts, connecting computers, workflows, running teams, questions, offline behaviour, a CLI and agent-tools reference, states, limits, security, privacy, troubleshooting and this FAQ.
llms.txtand per-page Markdown, for coding agents that read docs.
12 October 2026
First public release
Sculptured's core is open source under Apache-2.0.
- Design a team on a board: agents, groups, inputs and "hands work to" arrows, with things to fix pinned to the part they are about.
- Run it and watch it: each agent's live state, transcript, tool calls, questions and handoffs, with reported, observed and verified evidence kept apart.
- Steer it: send guidance, interrupt, soft stop, hard stop, restart, and resume after a restart.
- Claude Code, Codex and Practice agents, with token limits, output budgets, plan-usage holds and cost estimates that name their source.
- Reusable groups, Community blocks and the CS2680 pattern catalog.
- Hosted beta (invite-only): accounts with Google, GitHub or password sign-in, connected computers, an admin console and a What we record page listing everything recorded.