An AI coding task can stop for several reasons: a test fails, the agent needs approval, or you have to switch projects. Starting another terminal is easy. Remembering what each terminal was doing is the harder part.
I maintain Termexo, an MIT-licensed Windows workbench for organizing AI coding terminals. Its direction is to keep projects, agents, model settings and session context together, with browser access to the same running desktop when needed. This post describes a workflow to try, including the parts that still need a person.
Give every terminal a project and a job
Start with one repository and two tasks: implementing a change and checking it. Write an acceptance condition before starting either task. For a login form, that could be: an invalid credential response shows a useful message, and the corresponding test passes.
Termexo groups project directories, terminal layouts and related settings into workspaces. It supports Claude Code, Codex CLI, OpenCode and Antigravity, presenting their real terminals. The task board helps connect the work to goals and acceptance.
Repository screenshot with demonstration data. A task reaching a completed state still needs review.
When two agents work concurrently, define which files each may change and which task depends on the other. A shared workbench does not automatically resolve conflicting edits. Separate branches or worktrees can help when appropriate; the developer remains responsible for integration.
Treat waiting as something you need to see
A running process is not necessarily making progress. It may be waiting for input or approval. Status indicators and notifications help expose those moments so you can return to the relevant terminal and make the decision.
For a small trial, notice whether you can identify the waiting task without opening every terminal. That is a more useful evaluation than counting how many agents are open.
Leave a handoff that another session can use
Native session search and recovery help reopen supported agent sessions. When changing tools or handing work to another person, use a reviewed handoff with four parts:
- Goal: what behavior should change, and how will it be accepted?
- Progress: what changed, and where is the code?
- Verification: which checks ran, what passed, and what remains unverified?
- Next step: one concrete action, including unresolved constraints.
For example: “The login error message is implemented. The unit test passes. Browser behavior has not been checked. Next, verify keyboard submission and add the missing test if it fails.”
Termexo's handoff packages help carry this background. They are not a complete transfer of an agent's private context, and the receiving environment still needs the code, CLI and relevant configuration.
Reach the running desktop from another device
Browser access can be useful when a long task needs a short response while you are away from your desk. Trusted local networks or controlled VPNs can use direct access; cross-network access can use a reachable self-hosted HTTPS relay.
Project interface imagery illustrating access to the desktop from a phone.
The Windows host must stay on, with Termexo running and connected. The terminal processes execute there. Plan for sleep settings, connectivity and access credentials before relying on the remote workflow.
Where this fits today
For an individual developer, the use case is maintaining several projects without repeatedly reconstructing their state. Within an engineering team, the current fit is each member organizing their own workbench and producing clearer handoffs.
Shared workspaces, fine-grained permissions, centralized audit and fuller automatic orchestration remain longer-term plans. Local-first describes where the workbench and its state run; sending code to an external model still depends on the agent, endpoint and model configuration.
Try a complete cycle: define a task, implement it, check it, resume a session and review a handoff. Add remote access only if it solves a real interruption in your day.
Source and releases: gemron/Termexo. Setup and remote access: project guide. Agent CLIs, accounts and model services require their own setup.
Written by the project maintainer with AI assistance. Images come from the project and include demonstration content.
Top comments (0)