Agent task workbench
A fresh machine for
every coding task
Tezgah gives each task its own disposable container — a real, interactive Claude Code session, fully provisioned with your database, services, and a live preview URL. Run them in parallel. Nothing lives on your laptop.
Subscription-billed · never API credits
[01] The workbench
Everything a task needs to actually ship
A terminal and a chat box aren't a development environment. Tezgah wraps each coding agent in a real, provisioned machine — so it can run your app, see what's broken, and work without ever touching your laptop.
// 001
Disposable containers
Every task is a thin Incus container — its own repo checkout, branch, and services. Throw it away anytime; the git branch, the transcript, and the database all survive.
// 002
Provisioned end-to-end
Checkout, database, migrations, running services, and a live HTTPS preview — all stood up in under a minute. The agent's .env looks exactly like a normal local dev setup.
// 003
Live preview URL
Each task gets its own *.tezgah.app over HTTPS. Share work-in-progress, receive Stripe and GitHub webhooks, and click around the running app while you review the diff.
// 004
Fully parallel
One task = one container = one tmux = one branch. Spin up eight at once and nothing collides — separate ports, separate databases, separate redis indexes, leased up front.
// 005
Subscription-billed
Every agent session is a normal Claude Code session on your Max subscription — full TUI, your CLAUDE.md, your MCP. A runtime wrapper makes accidental API-credit billing impossible.
// 006
Archive & wake
Idle tasks stop and drop to zero RAM and CPU — disk, DB, and URL all kept. Wake them in seconds with the conversation intact, resumed straight from where the agent left off.
[02] Beyond the terminal
Everything your agent can't do alone
A coding agent can write code, but it can't run your stack, hold a preview URL, or survive a reboot. Tezgah is the harness that turns a chat session into a full development workflow.
Attach from any terminal
ssh, mosh, or the tezgah CLI — every surface attaches to the same live tmux session. Multi-attach is native; a teammate can watch the agent work in real time.
Coordinate across repos
Each task carries its own project and contract. Run tasks against different codebases side by side — different languages, different service shapes, no cross-talk.
Resume exactly where it stopped
A session supervisor baked into the image rebuilds the tmux layout and resumes the agent on every container start. Reboots, wakes, and crashes are one recovery path.
Nothing on your machine
Tasks run on a dedicated host, not your laptop. Commits and PRs come from you, billed to your subscription — but the work, the disk, and the risk stay off your box.
justmark
upcoach
[03] The lifecycle
From one command to a live, reviewable task
tezgah new runs the whole pipeline as a reconciling state machine. Archive is free and reversible; destroy is the only destructive verb — always explicit, never automatic.
provisioning
Leases a port, redis index, DB and branch — then clones, renders .env, installs, migrates and seeds.
ready
Site answers /up over HTTPS, the agent is briefed and waiting in tmux window 0. Under a minute.
working
The agent edits, runs tests against the real per-task DB, and hits its own live site. Attach to steer.
pushed
Branch on GitHub, PR opened as you — not a bot. The preview keeps running while you review.
archived
Stopped after idle days — zero RAM, everything kept. tezgah wake brings it back, conversation intact.
Every step writes an event row. A failure parks the task on the exact step → tezgah retry reconverges from there. Ports, redis indexes and DB names are transactional leases, taken before any side effect.
Put every agent at its own bench
One container per task. Fully parallel. Fully disposable. Billed to your subscription, not API credits.