Concepts

RexOS is designed for long-running work where “one prompt” isn’t enough.

Workspace

Most RexOS commands operate on a workspace directory:

• tools are sandboxed to it (filesystem + shell working directory)
• the harness stores durable artifacts there

Memory (SQLite)

RexOS persists:

• sessions
• chat messages
• small key/value state

in ~/.rexos/rexos.db, so later runs can resume with context.

Tools (sandboxed)

The agent can call tools such as:

• fs_read / fs_write (workspace-only, blocks .. traversal)
• shell (workspace-only)
• web_fetch (SSRF-protected by default)
• browser_* (headless browser automation via a Python Playwright bridge)

Browser tools prerequisites

browser_* tools require Python + Playwright:

python3 -m pip install playwright
python3 -m playwright install chromium

Model routing

RexOS classifies runs into a task kind:

• planning
• coding
• summary

Each kind can route to a different provider/model pair.

Harness (durable long tasks)

The harness adds a workflow on top:

1. initialize a workspace with durable artifacts
2. run incremental sessions
3. verify via init.sh / init.ps1
4. checkpoint via git commits