Solutions · Coding agents

Coding agents need somewhere to work.

The agent loop is excellent. What it doesn’t come with is a place to run code, files that survive a handoff, a database to read, or a trace of what it did. sys9 is that substrate — atomic services you compose under Claude Code, Codex, OpenClaw, Cursor, or OpenCode without touching the loop.

Request access See run9 →

Wrap, don’t rewriteUse-first, claim-laterPay only when running

what you compose

A workspace, assembled from four services.

Each does one job. Start with a sandbox; add files, a database, and traces as the agent needs them.

run9

A sandbox per task

Fork the whole OS — data and apt packages included — in ~50ms. Run a hundred in parallel, drive over SSH, pay only when running. Branch the whole stack to try a risky change.

run9 → drive9

Files that follow the agent

A shared FUSE filesystem across sandboxes, sessions, and handoffs — plus semantic grep (vector + BM25 in parallel) so the agent finds the right file, not just the named one.

drive9 → db9

A database from the terminal

Serverless Postgres with branching — fork the db, its data, and its files together for review. File storage built in, no S3, zero signup.

db9 → owl9

Traces, zero-config

Wrap the binary — no agent changes — and every session is captured: sub-agents auto-traced, human-readable transcripts, consistent across wherever the agent runs.

owl9 →

the shape

Fork a sandbox, run the agent, keep the trace.

One task, one forked sandbox. The agent works in isolation; owl9 captures it; drive9 finds the right file. Throw the sandbox away when it’s done — you pay only for the time it ran.

one task, one sandbox

Isolated by default. Disposable by design.

Every run gets its own forked OS, so a hundred agents on one repo never step on each other. Wrap the binary for traces; semantic-grep the shared filesystem for context. Nothing to provision, nothing to tear down.

~50ms fork up to 1M in parallel sub-agents auto-traced SSH-drivable

terminal

run9 fork my-repo --as task-41
  ✓ sandbox up in 312ms · pay only when running

OWL9_ENABLE=1 claude -p "fix the failing test"
  ✓ session captured · sub-agents auto-traced

drive9 fs grep "retry budget" /
  ✓ vector + BM25, in parallel

db9 branch my-app review
  ✓ db + data + files forked for review

works with your loop

Bring the agent. We bring the substrate.

sys9 sits underneath the runtime you already use — owl9 captures it by wrapping the binary, so there are no agent changes to make.

Claude CodeOpenAI CodexOpenClawCursorOpenCodeCline

The services behind this shape.

run9

use-first, claim-later

Give your coding agent a real workspace.

Hand an agent the sys9 skill, or install the CLI and fork your first sandbox in seconds. No signup to start.

Request access Running several at once? →