The harnesses

Choosing a harness

Escouade starts the harness on your machine, with your own account: install at least one and sign in to it once, as you would to use it in a terminal. Settings › Harness lists each one with its executable, its version and its state (see the harness settings).

As soon as two harnesses are installed:

• each project has its “Default harness” (Settings › Projects), Claude Code when unset;
• the “▾” button next to “New agent” offers “New agent on …” for each one;
• a ticket’s window has its “Harness” field, the project’s by default, which the autopilot follows;
• each agent’s card bears the mark of its harness, as soon as agents run on two harnesses.

An agent keeps its harness. When no account of its own can take it anymore, it can continue on another harness, with its conversation, its worktree and its branch.

Whatever the harness, the agent has the same conversation: the answer and the reasoning live, its tools under the same names (Bash, Edit with its diff, Read, Task…), the same permission, question and plan-to-approve cards, the “Plan” banner with its tasks and subagents, resuming after a restart and “Duplicate the conversation”. What a harness can’t do doesn’t show for its agents.

Comparison

Claude Code

Installation and account

Install Claude Code with its script (irm https://claude.ai/install.ps1 | iex on Windows, curl -fsSL https://claude.ai/install.sh | bash on macOS and Linux), then sign in with your Claude subscription or an API key. Escouade finds it in the PATH or in the usual places, even in WSL (see installation). Its accounts each have their folder (CLAUDE_CONFIG_DIR): see several accounts.

Models and modes

The models are those the Claude Code of each account lists in its /model, with their name, their efforts and whether each takes the “Auto” mode: a model one account has more (Mythos, say) shows in its agents’ Composer, a model it lacks does not. Escouade reads them at startup and when an account is added, without using anything, then at each agent start. The Kanban and the tickets offer those of the active accounts. An agent that moves to an account without its model takes the default one, and its conversation says so. Before the first reading, the list is Fable, Opus, Sonnet and Haiku. The five modes have their original meaning: “Auto” (a classifier approves safe actions), “Ask”, “Plan”, “Accept edits” and “Bypass” (see the permission modes). Claude Code’s permissions are what protects.

Cost, quota and view

A turn’s cost is the one Claude Code gives. The quota is your account’s 5-hour window and weekly window, in the status bar. The Claude Code view runs the agent in Claude Code’s own interface. It alone has remote control, background tasks and reading PDFs.

Codex

Installation and account

npm install -g @openai/codex, then codex login with your ChatGPT account. Escouade starts the native binary of the npm package (on Windows, the package’s codex.exe, not its launcher): Codex’s desktop app often installs an older version, which the PATH may find first. Each agent has its codex app-server. Its accounts each have their CODEX_HOME: see Codex accounts.

Models and modes

The models and their efforts are those Codex lists for your account. The modes go through Codex’s sandbox:

A request from Codex has three buttons: “Allow”, “For this session” and “Deny”. A message sent while it works enters its turn instead of waiting.

Cost, quota and view

Codex doesn’t give the cost: Escouade estimates it from the turn’s tokens and OpenAI’s public prices, and writes it “≈ … estimated”. The quota is your account’s windows (5 hours and 7 days, or a single 30-day one on the free plan). The “Codex view” runs the agent in Codex’s interface (see the terminal view).

What Codex doesn’t do

It doesn’t read PDFs (the button refuses them and says so) and offers no / commands in Escouade. A Codex agent doesn’t resume on another account: it waits for the reset, or continues on another harness. No remote control or background tasks.

OpenCode

Installation and account

npm install -g opencode-ai installs OpenCode 1, the only one Escouade drives: OpenCode 2 (@opencode/cli) is recognized and says “Not supported”. Connect a provider with opencode auth login (“Sign in…” in Settings › Harness runs it); Ollama’s local models work too. OpenCode has a single account in Escouade, its usual config. A single opencode serve serves all OpenCode agents, protected by a random password that never leaves Escouade; each agent is a session on it.

Models and modes

The “Model” menu lists the models of the connected providers, with a search, their provider and favorites (★) first, also set in Settings › Harness › OpenCode › “Models”. The effort is the model’s variant when it has one. Escouade always gives the model, so that OpenCode doesn’t pick one by itself.

OpenCode has no sandbox: its permission rules are what follows the mode, and an accepted command can do anything.

A request from OpenCode has “Allow”, “Always allow” (the card says the pattern OpenCode will keep for the session, like git add *) and “Deny”. A message sent during a turn waits for its end.

Cost, budget and view

A turn’s cost is the one OpenCode gives, at the provider’s price: 0 with a local model. OpenCode has no quota: the status bar shows its account’s cost for the day, and “Daily budget ($)”, in Settings › Accounts, pauses the autopilot until midnight once reached (see the OpenCode account). The “OpenCode view” plugs OpenCode’s interface into the agent’s session without stopping Escouade’s tracking (see the terminal view).

What OpenCode doesn’t do

No “Auto” mode (OpenCode has no reviewer), no sandbox, a single account, no attached PDFs, no / commands in Escouade, no remote control or background tasks.

Copilot CLI

Installation and account

npm install -g @github/copilot, then copilot login with a GitHub account that has Copilot (Free included). Each agent has its copilot --server --stdio, without the GitHub MCP server Copilot plugs in by itself, without access to the system’s temporary folder and without the GitHub tokens in its commands’ environment. Its accounts each have their COPILOT_HOME: see Copilot CLI accounts.

Models and modes

The models menu writes each one’s multiplier (×3), or its relative cost on a Free account; auto is never offered. The model and the mode change at once, even during a turn.

Copilot has no sandbox when Escouade drives it: Escouade’s rules answer the safe requests without a card.

A request from Copilot has “Allow”, “For this session” and “Deny”; no “Always”, which would write a lasting rule into Copilot’s config. A form from an MCP server becomes a card with fields. A message sent during a turn enters it at once.

Cost, credits and view

A turn’s cost comes from the AI credits of its calls (one credit is $0.01). The “credits” bar says what is used this month, until they come back on the 1st of the month; a session’s credit cap, once reached, becomes a card. The “Copilot CLI view” runs the agent in Copilot’s interface (see the terminal view).

What Copilot CLI doesn’t do

No “Always allow”, no assisted mode (experimental), no attached PDFs, no / commands in Escouade, no remote control or background tasks. A Copilot agent doesn’t resume on another account.

The MCP server

Whatever their harness, Escouade’s agents get its MCP server, each with its own token, when “Agents can drive Escouade” is on. The same switch declares it in each installed harness, so that Claude Code, Codex, OpenCode or Copilot CLI started outside Escouade, in a terminal or their app, drive Escouade: see the MCP server.

Written texts

A new agent’s name, the suggested commit messages and pull request descriptions are written by the harness of Settings › Application › “Model for written texts”: “Claude Haiku”, “Codex (GPT-6-Luna)”, an OpenCode model or “Copilot (GPT-5 mini)”, among the installed harnesses (see written texts).