AIOS Workspace

An open, agent-native operating system for an individual contributor — the workspace you work in day to day and from which you push selected output to a shared AIOS Team Brain.

AIOS Workspace gives you a clean folder structure, a set of governance conventions, validators that keep the repo honest, a growing library of dynamic multi-agent workflow harnesses that do real operational work — auditing a decision log, catching scope creep, turning meeting transcripts into decisions, synthesizing the week — with adversarial verification and rubric-gated self-correction so their output is trustworthy, and a sync client + local GUI for deciding what leaves your machine and pushing it to the brain.

AIOS is organized across three repositories, with distinct responsibilities:

• This repo — the individual workspace. One per person. You work here; you choose what to share. Nothing leaves until you aios push it.
• The standalone GUI. The local web GUI and desktop shell. It lives there, not here (AIO-612).
• The Team Brain. The one shared service that receives everyone's pushes and answers questions across the team. It is the only "team" layer.

It is designed to be cloned per person and shaped to your context, asked once at onboarding:

Are you a consultant working in a team for a client, or an employee working inside a company?

Your answer selects the spine that scaffolds — the same skeleton with a context skin (client/engagement framing, or internal role/OKR framing).

aios-workspace/
├── scaffold/        the workspace template (structure + rules + skills)
│   └── .claude/
│       ├── rules/    decision-log · frontmatter · publishing · hours · interlinking
│       ├── skills/   dynamic-workflow harnesses (decision-audit, scope-creep,
│       │             transcript-decisions, weekly-synthesis, aios-sync)
│       ├── rubrics/  checkable criteria for rubric-gated self-correction
│       └── memory/   instincts + incidents (cross-session learning)
├── validation/      OGR01–OGR15 validators (workspace hygiene · scaffold + runtime
│                    contracts · advisory scorecards); validate-all.sh runs them all
├── hooks/           Claude Code guards (secrets · access · frontmatter · sync nudge)
│                    shipped into every scaffolded workspace + registered in
│                    .claude/settings.json, so the PreToolUse guard fires there too
├── scripts/         scaffold-project.sh · aios.mjs (Team Brain sync CLI) · leak-gate.sh
├── packages/        @aiosbrain/foundation (npm workspace) — the shared hub modules,
│                    published to npm; scripts/ paths re-export it as thin shims
├── examples/        a fully synthetic sample to demo + test the harnesses
└── docs/            architecture · feature-set · workflows · brain-api (sync contract)

Install the current stable workspace

For onboarding, clone the immutable public release rather than the moving main branch. Resolve the newest tag instead of typing a version, so this never points at a superseded release:

REPO=https://github.com/aiosbrain/aios-workspace.git
LATEST=$(git ls-remote --tags --refs --sort=-v:refname "$REPO" 'v*' \
  | head -1 | sed 's#.*refs/tags/##')

git clone --branch "$LATEST" --depth 1 "$REPO" aios-toolkit
cd aios-toolkit

Then use scripts/scaffold-project.sh as shown below. A plain git clone checks out main; that is the development surface for contributors and can contain merged work that has not passed release promotion yet.

The CLI is also distributed on npm under the stable latest dist-tag:

npm install -g @aiosbrain/aios@latest
aios --help

The npm package carries the sync client, workspace scaffold (scripts/scaffold-project.sh), the OGR validators (validation/), the governance hooks, and the pinned contracts (docs/brain-api.md, docs/ENGINEERING-CONSTITUTION.md). For the current onboarding flow, keep the tagged toolkit checkout as the source used to scaffold and update workspaces; the global package is a convenient CLI distribution, not a replacement for that checkout. The unscoped aios package on npm is unrelated.

Release-tag checkouts do not follow main. When a later release is published, fetch and check out that exact tag, then apply it deliberately from your workspace with aios update --from /path/to/aios-toolkit --no-pull. Release notes will name the tag and any migration steps.

Quickstart

Want to actually drive this? docs/GUIDE.md — The AIOS operating manual — is the one task-oriented walkthrough of the whole surface, organized around your day: the daily brief, the asks queue + attention mode, the weekly loop, syncing to the brain, measuring yourself (AM), gating specs, and the agent pipeline. Start there.

New user? docs/GETTING-STARTED.md is the step-by-step stable-release path to your first aios push live on the Team Brain.

Scaffold your workspace — pick the context that matches how you work:

# Consultant working in a team for a client:
scripts/scaffold-project.sh --context consultant \
  --slug acme-workspace --stakeholder "Acme Corp" \
  --owner alex --team "alex,sam,jordan" \
  --org your-github-org \
  --output ~/Projects/acme-workspace

# Employee working inside a company:
scripts/scaffold-project.sh --context employee \
  --slug alex-workspace --stakeholder "Acme Inc" \
  --owner alex --team "alex,sam,jordan" \
  --org your-github-org \
  --output ~/Projects/alex-workspace

(Legacy flags still work: --profile engagement maps to --context consultant, and --lead/--captain/--client/--members are accepted as aliases.)

Validate any workspace — from anywhere, including from inside the workspace itself:

aios validate                                  # the workspace you are standing in
aios validate ~/Projects/acme-workspace        # or name one
aios validate --critical                       # OGR03 secrets scan only

aios validate resolves the validators inside whichever toolkit is installed, so a global npm install does not need $(npm root --global)/@aiosbrain/aios/validation/validate-all.sh — a path that differs per Node install manager (nvm, fnm, Homebrew, Windows). The underlying script is still directly runnable from a toolkit checkout, pointed at a workspace — a scaffolded workspace has no validation/validate-all.sh of its own:

validation/validate-all.sh ~/Projects/acme-workspace   # run from the toolkit checkout

Run a harness (via Claude Code's Workflow tool) against the included sample:

Workflow({
  scriptPath: "scaffold/.claude/skills/decision-audit/decision-audit.workflow.js",
  args: { repoPath: "examples/sample-engagement", runDate: "2026-06-05" }
})

Prefer to just see results? examples/sample-output/ has real harness output on the synthetic sample — a decision-log audit (9 verified findings, false positives filtered out), a scope-creep register, and decisions extracted from meeting transcripts.

Connect to a Team Brain (optional — everything works offline without one):

cp .env.example .env          # add your AIOS_API_KEY
git config aios.member alex   # your identity
aios status                   # what would sync, what's blocked, why
aios push                     # push team- and outward-tier content
aios query "what's blocking sprint 2?"

Chat with your repo in a browser instead of the terminal — the local cockpit:

git clone https://github.com/aiosbrain/aios-workspace-gui.git
cd aios-workspace-gui && npm install && npm run build:client
node gui/server/index.mjs --repo ~/Projects/acme-workspace --toolkit-dir /path/to/aios-workspace

Repo topology note: the GUI and desktop shell live in aiosbrain/aios-workspace-gui (history filtered from this repo). gui/ and src-tauri/ were deleted from this repo in AIO-612 — that repo is the only copy. The GUI locates a toolkit checkout via docs/gui-toolkit-contract.md: --toolkit-dir flag → AIOS_TOOLKIT_DIR env → an actionable error. The pre-split relative fallback (gui/server/../../) still exists in the GUI repo's resolver but can no longer reach a toolkit — nothing sits above it any more — so one of the two must be set. It builds against the published @aiosbrain/foundation package plus the aios CLI — never toolkit internals.

The cockpit is more than a chat box:

• Model picker — switch between Sonnet 4.6 (the fast, cheap default) and Opus 4.8 live, mid-session, with no reconnect.
• Chats — a resumable chat-history sidebar; each conversation is saved and reopened where you left off (+ New chat to start fresh).
• Context (est.) meter — an approximate read of how much of the model's window the last turn used, so you can see when a chat is getting heavy.
• Markdown rendering — assistant replies render as GitHub-flavored markdown (tables, lists, code).
• Settings → Personality — pick the agent's voice (AIOS · Analyst · Coach · Operator); it's a style layer over the system prompt and never overrides your rules, CLAUDE.md, or skills.
• Skills — install official, Apache-2.0 Anthropic skills (vendored and hash-locked) into .claude/skills/ with one click; document skills (Word, Excel, PowerPoint, PDF) are surfaced as pointers to Enable in Claude.

First-run onboarding can draft your profile from a link — paste a company or profile URL and the agent reads it with Firecrawl, then drafts your .claude/CLAUDE.md for you to confirm (connect Firecrawl in Integrations first).

Not sure what else is available? npm run help lists every script grouped by category (Core / Dev / Build / Internal), instead of the undifferentiated dump a bare npm run prints.

The numbered spine

Every workspace uses the same six-folder pipeline, raw → refined. The 0-context and 4-shared folders take a context skin; the rest are identical either way:

Access tiers

You choose what leaves your machine. Content is tagged with a friendly tier label that maps to the engine's canonical tier:

Default-deny: untagged content and anything private/admin never syncs. Promotion is always a deliberate aios push.

The daily driver — loop, asks, mode, specs

Beyond scaffolding and sync, the workspace runs an operator loop and a human-operating layer you drive from the CLI — all offline, all local-only until you push:

• Daily brief — aios loop daily: what changed, what's blocked, what you owe today, with your own escalation queue on top.
• Asks queue — aios asks list/add/resolve/drain: a non-blocking escalation inbox that session hooks fill for you (idle → blocker, an assistant question → decision, a completion → fyi).
• Attention mode — aios mode deep-work | orchestration: one command to silence or restore the local end-of-turn ping (touches only preferredNotifChannel; mobile push untouched).
• Steering decisions — aios decisions list/outcome/export: a local, admin-tier corpus of the AskUserQuestion / plan-approval moments (this is not the Team Brain decision log — see the guide's disambiguation box).
• Weekly loop — aios loop collect → weekly → verify → writeback: a verified, approval-gated closeout that proves every shareable claim before it can sync.
• Measure yourself — aios analyze (agentic maturity (AM) + attention + spend), aios time (agent-runtime time tracking).
• Spec gate — aios spec eval | fix: grade a spec against the pick-up-able-issue bar before a builder touches it.
• Agent pipeline — aios relay | build | ship: an Opus builder reviewed by an independent model, gated at plan + merge. See docs/agent-build.md.
• Repo hardening — aios assess-codebase, aios rails missing/suggest/apply: score any repo's agent-readiness and stand up a safe permission allowlist.

Every one of these is walked through, with real output and diagrams, in docs/GUIDE.md.

Skills & integrations

Every workspace ships a generated skills catalog (.claude/skills/INDEX.md) and an integrations catalog (.claude/INTEGRATIONS.md, from .claude/integrations.json) — so you can see what the workspace can do and connect to (Slack, Jira, Notion, Linear, GitHub, Gmail, Granola, Confluence, Mattermost, Toggl). To wire an integration, copy its server from scaffold/.mcp.example.json into scaffold/.mcp.json, set the env vars, and run npm run gen:catalog. Setup notes: docs/integrations.md.

Skills are shareable: aios push skill <name> publishes to the brain; aios pull skill <name> + aios install-skill <name> adopt one (install is always explicit — pulled skills never auto-activate).

Terminology

The toolkit started life in consulting; both vocabularies are accepted everywhere (validators, hooks, harnesses) and existing clones never need to rename:

See docs/GUIDE.md for the task-oriented operating manual (the whole CLI, organized around your day), docs/architecture.md for the hub-and-spoke model and access tiers, docs/feature-set.md for the full feature set, docs/workflows.md for the harness design study that shaped the skills, docs/byoa.md for Bring Your Own Agent (choose your runtime — Claude Code, Hermes, OpenClaw, Codex…), and docs/brain-api.md for the Team Brain sync contract.

Status

Early and open. The clean core (structure, rules, validators, guard, harnesses), the rubric-gated self-correction layer, the Team Brain sync client, and the local GUI are here; more harnesses and integration adapters are on the roadmap. Contributions welcome — see CONTRIBUTING.md.

License

AGPL-3.0-only — see LICENSE and LICENSING.md.

Running a workspace is unrestricted; the AGPL places no obligation on internal use. Two directories are Apache-2.0 deliberately: packages/integration-sdk/ (the integration contracts) and .harness/ (which aios repo-bootstrap stamps into your repository) — so writing an integration, or bootstrapping your own repo, never licenses your code. Prior releases remain MIT, preserved in LICENSE-MIT. If your organization's policy bans AGPL, there is a free commercial license for internal use — email cn@fluora.ai.