AI-generated concept artwork, not a product screenshot. Chinese labels describe the local workbench, terminals, tasks and settings.
A coding agent can change a repository while remaining unaware of the test running in another terminal or the tasks already recorded in your workbench. Copying that information into a prompt works, but it creates another manual step whenever the development session changes.
I maintain Termexo, an MIT-licensed Windows workbench for coding terminals. Version 0.10.10 exposes 19 local MCP tools and adds automatic connections for new Claude Code, Codex, OpenCode, Grok Build and Antigravity terminals launched by the application. This post focuses on a small, observable trial and the distinction between connecting a tool and completing a task.
Start with a query you can verify
Install the Windows EXE or MSI, create a workspace for a local repository, and open Settings → AI Control (MCP). Enable the server, apply the selected access scopes, and keep automatic Agent connections enabled. Start a new agent terminal; existing processes must restart to load the new configuration.
For the first request, ask:
Call workspace_list from termexo-desktop and list the workspaces. Do not create tasks, write terminal input or change settings yet.
Check for an actual tool call and the expected workspace. Keep the Termexo desktop open for the connection. Claude attach sessions refer to an existing process, so that original process also needs to restart.
Cover frame from a real v0.10.6 desktop recording, illustrating side-by-side terminals. It does not show the new v0.10.10 MCP settings.
Add one operation at a time
The tool surface covers workspace and terminal queries, terminal creation and input, task creation and lifecycle operations, and selected settings. Terminal, task and settings scopes are controlled separately. Settings access needs separate enabling. Reading configuration excludes credentials.
After the initial query, identify one test terminal and ask the agent to read its output. Then try creating a draft task with a project, goal and acceptance criteria you can inspect. This sequence makes it easier to distinguish a wrong target from a failed connection.
Task execution currently supports Claude Code, Codex, OpenCode and Grok Build. Antigravity can receive the MCP connection in its terminal, but cannot execute task-board tasks.
A successful terminal_write means input was delivered. A successful task_execute establishes the task/terminal launch binding. Neither proves the requested code works: read the output, inspect task state and review the resulting files before accepting the work.
Preserve configuration, and make removal possible
The server is disabled initially and listens only on loopback; the default address is http://127.0.0.1:7421/mcp. Automatically connected terminals use Termexo's stdio bridge. For each request, the bridge reads the current port and system credential; the token is absent from generated agent configuration and command lines. Client-side tool approval still applies.
Diagram of the implemented connection path, not application UI.
Claude Code, Codex and OpenCode use launch-specific configuration that preserves existing profiles or other services. OpenCode V1 and V2 need different configuration structures. Grok Build and Antigravity use user-level configuration, so clients launched outside Termexo may also see the entry. Disabling the server or automatic connection and applying settings removes Termexo-owned additions while preserving other services. Invalid configuration and an existing user-created entry with the same name produce errors.
External clients can use the JSON or TOML generated in Settings if they support Streamable HTTP and Bearer authentication. This requires their own configuration.
What the integration test actually established
The official MCP SDK completed the stdio handshake against the released Windows application, discovered 19 tools and queried an isolated desktop workspace. An automatically connected Codex task terminal made a real workspace_list call and received the test workspace.
Grok completed handshake and discovery; an isolated OpenCode 2.0.15 service confirmed connection; Antigravity recognized an enabled entry; Rust tests cover Claude configuration merging. These checks have different depths, and are not five equivalent end-to-end model tests.
The Codex task trial required an extra carriage return through terminal_write. Fully automatic prompt submission still needs further verification. The live test report records these boundaries.
Keep the development workflow accountable
Concurrent agents can still edit the same files. Separate directories or Git worktrees provide file isolation; the MCP connection does not resolve conflicts. Agent CLIs, accounts and model services need their own setup, and the project does not include model subscriptions.
If you use several coding agents on Windows, begin with one workspace query, one terminal read and one draft task. Evaluate those concrete operations before extending the workflow.
Source and release · Setup guide
Published by the project maintainer. The article was prepared with AI assistance; concept artwork, recording imagery and the technical diagram are labeled separately.
Top comments (0)