~/shelltab — docs/ai-agents

AI Agents

Run any AI agent inside a drive and watch it work in real time. Every command it runs is attributed, recorded, and reviewable — so you're always in control of what happened and why.

ai-agents-overview.mp4

How agents run in ShellTab

An agent in ShellTab is a process running inside your drive — the same persistent Linux environment your team uses for terminals. The agent has full access to the filesystem, network, and installed tools. It can clone repos, run tests, write files, and execute arbitrary shell commands, just like a human teammate would.

What makes ShellTab different is visibility. Every command the agent runs appears in a dedicated agent tile on the canvas with the agent's identity stamped on each output block. Your team can watch the agent work in real time, see exactly what it's doing, and step in at any point.

Agents are attributed as a named service identity (e.g. claude-code). Their commands appear alongside human commands in the session recording and can be reviewed, replayed, or audited after the fact.

Supported agents

Claude Code

Anthropic's coding agent. Start a Claude Code session directly from the ShellTab toolbar. It opens in its own tile with a dedicated composer for sending prompts.

OpenAI Codex

Run Codex via the OpenAI CLI inside a terminal tile. ShellTab captures and attributes every command it runs automatically.

Custom agents

Any process that runs shell commands works. If your agent uses the shell, ShellTab captures it. No SDK integration required.

Shell scripts

Long-running automation scripts are attributed to the user who started them. Useful for deployments, migrations, and scheduled jobs you want to audit later.

Starting a Claude Code session

  1. 01 Open the toolbar

    Click the + button in the canvas toolbar and select Claude Code. A new agent tile appears on the canvas.

  2. 02 Write a prompt

    Type your task in the composer at the bottom of the agent tile. Be as specific or as open-ended as you like — Claude Code will ask clarifying questions if it needs to.

  3. 03 Watch it work

    The tile shows the agent's reasoning and every shell command it runs. Output streams in real time. Other teammates can watch alongside you.

  4. 04 Step in or redirect

    Click anywhere in the terminal area of the tile to take control. Type a correction, run a follow-up command, or send a new prompt to redirect the agent.

ai-agents-claude-code-session.mp4

Reviewing agent work

After an agent session ends — or while it's still running — you can open the Session Recording tile to replay everything that happened. The timeline shows each command in sequence, annotated with the agent's identity and the timestamp.

For code changes, ShellTab's review pipeline analyses what the agent wrote: which files it touched, what commands it ran in each phase of the task, and whether its work drifted from the original prompt. Findings are surfaced as a structured review in the Review tile.

Safety & oversight

ShellTab doesn't gate agent access — agents have the same filesystem and network permissions as a human user in the same drive. The safety model is transparency and auditability rather than sandboxing. You can see everything the agent does as it happens, and nothing is hidden after the fact.

If your use case requires a restricted environment, create a separate drive for agent-only workloads. That drive can have a minimal toolset and no access to production credentials.

What's next