Install · Claude Code · Forge shell · Commands · Dashboard · Docs

---

Why 8sync

Rust core, one file	A single static binary: 4.2 MB on Windows, 5.4 MB on linux-x86_64 (musl), no runtime to install. A size gate in CI fails any release that grows without an attributed reason.
One line to a full AI workstation	The installer drops the binary, then sets up gh, omp, codegraph, Forge with zsh, Claude Code with the 8sync kit, 50+ skills, MCP servers and global agent rules. Every step is idempotent and logged; a failed tool is reported and the rest still installs.
Claude Code that finishes the job	8sync ccc . <name> opens a named Claude Code session per repo. The kit adds subagents, /ship, a background compile check after every edit, every 8sync skill, the team's Usage Policy rules, and two gates: one keeps Claude working while it has unfinished work, the other sends questions it can answer itself back to it. Details.
Claude, the legitimate way	Claude Code signs in with Anthropic's own login; Forge and omp use your own API key on the official API. No scraped web sessions, no shared accounts. Details.
A shell that is also an agent	8sync forge <name> opens a fast zsh bound to a named Forge conversation: type commands as usual, prefix with : to ask Claude.
Agents that know the code	Code graph indexing (codegraph, codebase-memory-mcp, Serena), per-project memory in su-code/, and always-on engineering skills, so agents query structure instead of grepping blindly.
Same commands everywhere	Linux, macOS and Windows. On Windows there is no WSL requirement and no admin rights needed. zsh is provisioned privately next to Git for Windows.
Updates itself, safely	Every command checks for a new release in the background (throttled to once per 6 h). 8sync up installs it with a sha256 check, even while sessions keep the old binary open; 8sync ccc installs it on launch and re-runs on the new binary.

flowchart LR
  U([you, in any terminal]) --> X[8sync · Rust]
  X --> O[omp agent<br/>8sync . / 8sync ai]
  X --> K[Claude Code + kit<br/>8sync ccc . name]
  X --> F[Forge + zsh<br/>8sync forge name]
  X --> H[harness<br/>skills · rules · MCP · hooks]
  X --> M[(su-code/ memory<br/>STATE · KNOWLEDGE · DECISIONS)]
  O --> C[Claude via Anthropic]
  K --> C
  F --> C
  H --> G[code graph<br/>codegraph · cbm · serena]
  O --> G
  K --> G
  F --> G

Loading

---

Install

curl -fsSL https://8-sync-dev.github.io/su-code/install | sh       # Linux / macOS

irm https://8-sync-dev.github.io/su-code/install.ps1 | iex         # Windows PowerShell

You don't need git, Rust or admin rights. The installer verifies the binary's sha256 against the release, puts it on your PATH, then runs 8sync setup --no-profile:

Step	What you get
gh	GitHub CLI. Native package manager or winget; on Debian/Ubuntu, the official release tarball into ~/.local/bin (no sudo).
omp	The coding agent (oh-my-pi), from its official installer.
codegraph	Semantic code index used by every agent.
Forge + zsh	Forge from its GitHub release. zsh comes from the system; on Windows it is a private MSYS2 zsh (sha256-pinned); on Linux without sudo, a static portable build.
Claude Code + kit	Claude Code from Anthropic's official installer, then the 8sync kit into ~/.claude (below).
harness global	50+ skills, APPEND_SYSTEM.md agent rules, MCP servers (codebase-memory-mcp, headroom, serena), hooks, context compaction at 50%.
PATH	~/.local/bin, ~/.cargo/bin, ~/.bun/bin, ~/.encore/bin wired into zsh/bash/fish.

A real run of the v0.63.0 one-liner on a fresh Ubuntu account without sudo:

Options: SUSYNC_NO_SETUP=1 (binary only) · SUSYNC_VERSION=v0.63.0 (pin) · SUSYNC_BIN_DIR=~/bin · sh -s -- --uninstall. Run it again any time, or 8sync up. Fallback URL: https://raw.githubusercontent.com/8-Sync-Dev/su-code/main/install.sh.

CI runs these exact one-liners on Ubuntu (x64 and arm64), macOS (Apple silicon and Intel) and Windows after every release and every week (install-smoke); a run fails unless the installed version is the latest and setup reports "all steps succeeded".

Pro terminal (opt-in, any OS): 8sync setup --profile terminal installs:

• helix with a tuned config, plus the JetBrainsMono Nerd Font. On Windows the font also becomes the Windows Terminal default if you have none.
• 8sync's zsh: two-line git-aware prompt, arrow-key completion menu, history suggestions, syntax highlighting.
• kitty glass theme, on Arch/Fedora only.

Then check everything with 8sync doctor.

---

Claude Code: 8sync ccc

cd <project>
8sync ccc . core               # Claude Code session "core" in this repo (create or resume)
8sync ccc . core -p "…"        # one-shot; any flag after the name goes to claude
8sync ccc .                    # the session you used last
8sync ccc . ls | new <n> | rm <n>
8sync ccc kit                  # (re)install the kit; `8sync setup` already did
8sync ccc status               # claude, login, kit, sessions

Each name maps to a fixed conversation id per repo, so 8sync ccc . core tomorrow continues today's thread, next to 8sync forge core and 8sync . core for the other agents.

The kit lives in ~/.claude, so it works in every repo and for plain claude too. Every 8sync ccc launch also keeps things current, without network on the launch path: a newer 8sync release is installed and the command re-runs on it, the kit and bundled skills refresh after an update, and the repo gets the 8sync harness (memory, AGENTS.md/CLAUDE.md blocks, .claude/skills → su-code/skills). SUSYNC_NO_AUTO_UPDATE=1 keeps the installed binary.

Part	What it does
Org rules	The 8 Sync rules for Anthropic's Usage Policy and terms (8sync-anthropic-rules.md) are imported into every session ahead of other instructions.
Subagents	explore (Haiku) · task · quick-task · reviewer · designer. Reviewer and designer use your session model; the read-only ones keep the code-graph MCP tools.
Skills	/ship <task>: explore → task list → implement with build/test → reviewer → CHANGELOG + commit, never pushes. /aup-audit: Usage Policy + secret check before you publish. deep-research: cited research brief. Plus every 8sync skill: bundled ones linked to the current ~/.omp/skills copy, project ones through .claude/skills; on-demand skills are listed by name only to keep the prefix small.
Verify hook	After every Edit/Write, runs cargo check, the package's typecheck script (or tsc --noEmit) or go vet from the edited file's nearest manifest in the background; only a failure comes back to Claude. The latest edit wins, so a burst of edits costs one check. Before a turn ends, the build must pass (blocked once per broken state). Native Rust, works on Windows.
Guard hook	Before Bash/Edit/Write: no writing .env/keys/*.session/*.pkl, no real-looking API key in code or commands, no git add -A that would sweep up a secret, no code routing a Claude subscription into another client. About 23 ms per call.
Stop gate	When Claude tries to end a turn, Haiku reads its last message. Planned steps not done, edits not verified, failing checks, NOT MERGEABLE, or a question it could answer itself: the turn goes on.
Ask gate	Before Claude interrupts you with a question, Sonnet checks it. Only credentials, destructive or paid actions, and product decisions reach you; everything else goes back with "pick the conventional option and continue".
Defaults	Only where you haven't set them: auto permission mode, medium effort, a task list on every model. Build/test/lint commands and git add skip the auto-mode classifier. Deny rules for reading/editing .env, keys, *.session, *.pkl and for force-push; omp's codebase-memory-mcp + serena registered.

A real run with the kit. The first edit broke the build, the verify hook fed the compiler error back, and the stop gate refused "Next I will add tests":

Gate accuracy on labelled cases (the exact hook prompts sent to the model):

Gate	Model	Correct
Stop	Haiku 4.5	18 / 18 (6 cases × 3 runs)
Ask	Sonnet 5	6 / 6
Ask	Haiku 4.5	4 / 6, so not used

Your edits win: a kit file you changed is kept (8sync ccc kit --force overwrites it), and settings.json is merged, never replaced.

Limits, so you know what you are getting:

• The stop gate judges Claude's last message, not the whole transcript. Claude Code stops after 8 forced continuations in a row.
• AskUserQuestion does not exist in claude -p, so the ask gate only acts in interactive sessions. If its model id is ever retired, the hook fails open and questions go through.
• The verify hook checks the whole package, not only the edited file. It runs in the background, so a slow typecheck delays only the end of the turn (timeout 600 s).
• medium effort is the kit default because it halved the time on the bench below with the same hidden-test result; switch a session to /effort high for hard reasoning.

Speed next to omp (measured)

A harder task: an expression evaluator with three bugs and missing features across four files, 11 hidden tests, Sonnet 5 with a Console key, a fresh copy per run, up to four runs in parallel. Wall-clock per run:

Harness	Runs	Mean	Hidden tests
omp	155 · 124 · 136 s	138 s	11 · 10 · 11 of 11
8sync ccc, kit 0.64	281 · 216 s	249 s	11 · 11
8sync ccc, kit 0.65 at high effort	209 · 179 s	194 s	11 · 11
8sync ccc, kit 0.65 default (medium)	100 · 161 s	131 s	11 · 11

Where the time went: in kit 0.64 every edit waited for cargo check (2.5 s warm, 8.6 s cold in this repo), every build/test command waited for the auto-mode classifier (~0.65 s), and Sonnet 5 at high effort wrote 17.6-20.6k output tokens per run against 9.4k and 15.3k at medium. Run-to-run spread is large (100 vs 161 s at the same settings), so read this as "on par with omp", not "faster". Anthropic's parallel-tool-call prompt snippet did not change how often Claude batched calls, so the kit doesn't use it. Fast mode (/fast, Opus only, higher price) is the remaining lever and stays your call.

Tokens and quality next to omp (measured)

Same model (Sonnet 5, Console API key), same prompt, fresh copy of the repo per run, two runs per cell. Task A: implement a parser from a spec and fix three reported bugs (31 hidden tests). Task B: add save/load that survives hostile item names (20 hidden tests). Nobody saw the hidden tests.

Harness	Hidden tests	Input tokens A · B	Requests A · B	Cost A · B
Claude Code, no kit	31/31 · 20/20 (all runs)	523k · 347k	18 · 9.5	$0.24 · $0.14
8sync ccc, kit 0.63	31/31 · 20/20	555k · 344k	19 · 10.5	$0.26 · $0.16
8sync ccc, kit 0.64	31/31 · 20/20	490k · 272k	18.5 · 7.5	$0.29 · $0.19 ¹
omp	31/31 · 20/20	274k · 208k	8 · 7	$0.25 · $0.13

¹ First runs on a new config dir, so the prompt cache was cold (cache writes 39-48k vs 18-30k); input tokens are the cache-independent number.

What this says, and what it doesn't:

• On these tasks every harness got every hidden test right. They do not show Claude Code verifying better than omp; the kit's gates matter on long runs where a model stops early, which these short tasks never triggered.
• Claude Code processes 1.3-2× omp's input tokens because it makes about twice the requests: it mostly calls one tool per request, omp batches. The cost is close anyway, since repeated context is billed as cache reads at a tenth of the price. Telling Claude to batch calls in CLAUDE.md changed nothing (about 1 in 7 requests had several calls with or without it), so the kit doesn't.
• In an 8sync repo the per-request prefix is where the kit saves: in this repo it is 34.3k tokens for 8sync ccc 0.64, 41.0k for 0.63, 56.2k for omp. The difference is CLAUDE.md: 0.64 writes a 0.7k pointer block instead of a copy of the 8k skill list (importing AGENTS.md would cost 20.5k). 0.65 adds the always-loaded org rules and every skill by name: 41.1k in a fresh 8sync repo with 237 skills, against 57.3k if every description were listed.

---

Claude, the official way

8sync never logs into claude.ai for another tool and never proxies a consumer subscription.

• Claude Code (8sync ccc) is Anthropic's own CLI. Sign in with claude auth login (your Claude plan) or claude auth login --console (API billing).
• Forge and omp use an API key from console.anthropic.com against Anthropic's official API, under Anthropic's Commercial Terms:

8sync forge key sk-ant-...                 # Forge: stored in ~/.forge/.credentials.json, printed masked
export ANTHROPIC_API_KEY=sk-ant-...        # omp (`8sync .`, `8sync ai`): standard Anthropic env var
8sync forge key                            # list stored keys (masked)

Pick the model inside the tools: /model in Claude Code and omp, :model in a Forge shell. The keys stay on your machine, and 8sync never prints them in full.

8sync feynman auth-omp bridges omp's other logins into Feynman but never Claude; give Feynman a Console key (ANTHROPIC_API_KEY). 8sync harness claude-code with a non-Anthropic endpoint configures Claude Code only and leaves omp alone, and 8sync doctor warns when omp is signed in to Anthropic with a Claude plan instead of an API key. Other bridges (8sync codex-web) and gateways are opt-in tools under their provider's terms.

The team's rules live in assets/skills/aup-audit/8sync-anthropic-rules.md, ship inside the binary and load into every Claude Code session. A weekly workflow (legal-sync) snapshots Anthropic's own texts into docs/legal/anthropic/, commits any change and opens an issue to review the rules; with an ANTHROPIC_API_KEY repo secret, Claude Code drafts that update as a pull request.

---

Forge: the AI shell

cd <project>
8sync forge core          # zsh bound to the "core" conversation (create or resume)
8sync forge core --tui    # same conversation in Forge's full-screen TUI
8sync forge core -p "…"   # one-shot prompt
8sync forge ls | new <n> | rm <n> | status

Each name maps to a persistent Forge conversation per repo, so 8sync forge core tomorrow continues today's thread. In the shell, the prompt looks like this:

Prompt part	Shows
su-code/crates	Path relative to the repo root
on main	Git branch
core	Session name
FORGE 48.5k	Agent and tokens used in this conversation
claude-opus-5-5 MEDIUM	Model and reasoning effort
3.2s	Duration of the last command, when it took ≥ 2 s
❯	Green after success, red after a failure

Type	Does
: <prompt>	Ask Forge in this session's conversation
:<Tab>	Forge commands: :model, :provider, :agent, :new, :commit…
@<Tab>	Attach a file to the prompt
↑ / ↓	History search by what you've typed
→	Accept the grey suggestion
Tab	Completion menu (arrows to pick, Shift+Tab back)

Speed (Windows, measured in a pseudo-terminal):

• About 0.3 s from 8sync forge core to a usable prompt.
• The Forge segment renders asynchronously, so the prompt never waits for it.
• Branch and repo come from .git/HEAD directly, with no git process per prompt.

Your own ~/.zshrc loads first, and 8sync's managed files never overwrite it.

---

Commands

Daily loop

Command	What it does
8sync . · 8sync . <name>	Resume the project's omp session, or a named one. . new <name> --worktree gives a feature its own git worktree + branch; . merge <name> lands it locally.
8sync ai "…"	One-shot prompt; empty resumes the session. --model <alias> overrides the model.
8sync ccc . <name>	Claude Code session by name, with the kit (above).
8sync forge <name>	AI shell (above).
8sync find <kw>	ripgrep + fzf, opens file:line in your editor.
8sync run dev|build|test|fmt|lint	Project command from the detected stack.
8sync ship "msg"	add · commit · push · gh pr create.
8sync note "msg"	Append to su-code/NOTES.md; agents read it next session.
8sync feature new <slug>	Multi-phase feature plan with per-phase acceptance gates.

Harness

Command	What it does
8sync harness	One idempotent command per repo: skills, codegraph index, AGENTS.md force-load block, su-code/ memory.
8sync harness init · create	First bootstrap with managed .gitignore + gitleaks pre-commit hook · plus Cursor and Z.ai Code skill/rule fan-out.
8sync harness global [--sweep]	Apply rules, skills, MCP and hooks to every omp project on the machine.
8sync harness web	Local dashboard at http://127.0.0.1:8731 (see below).
8sync harness bench · toolstats · audit · eval	Context budget · code-intelligence vs grep usage · doc hygiene · quality suite.
8sync harness add-model · add-local-model · compaction	Register a remote model · serve a local GGUF via mistral.rs · set the compaction threshold.
8sync skill [add <url>[@ref] | update | gen]	Skill library; @ref pins a commit in skills.toml.

Lifecycle

Command	What it does
8sync setup [--profile <name>] [--dry-run]	Stage A harness + opt-in profiles: terminal, dev-stack, nvidia, warp, bluetooth.
8sync up [--to vX.Y.Z]	Self-update from GitHub Releases, sha256-verified. 8sync omp update updates omp.
8sync doctor · flow · help	Health check · workflow-ordered help · cheatsheet.

Tools for agents

Command	What it does
8sync shot · diff-img · pdf-img	Web page/file, git diff or PDF → PNG for vision models.
8sync gen-img "…"	Generate an image (gpt-image-2).
8sync locate <img> "<target>"	Visual grounding: boxes and click coordinates (LocateAnything-3B, non-commercial licence).

Machine (Linux desktop)

Command	What it does
8sync sec · vpn	WARP + ufw toggle · SoftEther + VPN Gate relays.
8sync bt · hz · lcd	Bluetooth repair · highest refresh rate per display · Lian Li screens.
8sync clean · theme · bg	Disk/RAM cleanup · kitty palette · kitty wallpaper.

Every verb has -h with an EXAMPLES block.

---

Dashboard

8sync harness web serves an axum + React app embedded in the binary. Every page reads real files, and most pages write changes straight back to them.

Group	Pages
Session	Live plan (su-code/STATE.md), token and compaction stats
Configure	Models per role, 50+ skills, memory files, rules
Runtime	Engines, code graph (call graph + Leiden clusters + caller tracing), MCP, submodules
Quality	Bench, readiness, team
Discover	One-click MCP servers and skills from the official registry, Smithery, Glama, mcp.so

---

Project memory

8sync . or 8sync harness seeds a memory every AI tool reads (omp, Claude Code, Cursor, Forge, aider):

<repo>/
├── AGENTS.md          anchor: force-loaded skills and rules
└── su-code/
    ├── STATE.md       the live plan, rewritten at every phase
    ├── KNOWLEDGE.md   validated lessons and recorded failures
    ├── DECISIONS.md   architecture decisions
    ├── PREFERENCES.md how you like things done
    ├── NOTES.md       `8sync note`
    └── skills/        project-local skills

---

Build from source

git clone https://github.com/8-Sync-Dev/su-code && cd su-code
bash scripts/bootstrap.sh            # rustup if missing → cargo build --release → ~/.local/bin/8sync

Topic	Details
Layout	crates/cli/src/main.rs routes; one file per verb in crates/cli/src/verbs/.
Embedded assets	Configs, skills and the zsh config in assets/, embedded with rust-embed.
Dashboard	Built from web/ by build.rs.
Binary size	bash scripts/size-report.sh attributes size; scripts/size-gate.sh enforces the ceiling in CI. --no-default-features drops the dashboard (3.1 MB).
Releases	A v* tag builds 5 targets and publishes them with digests.
Contributing	AGENTS.md is the guide for contributors and their agents.

---

License

MIT, see LICENSE. Built by 8 Sync Dev.