backends

Override the configured backend for a single invocation without editing config. Flags are global — they work on any ace subcommand.

$ ace --codex # launch Codex for this session only
$ ace --claude -- --continue # resume with Claude, passing --continue through
$ ace mcp --codex # check MCP against Codex without switching config

Schools, users, or projects can register additional backend entries via [[backends]]. A custom backend aliases a built-in (claude, codex, and opencode) and may override its launch cmd and env. The MCP, instructions file, and linked-folder behavior are inherited from the aliased kind.

# Override env on the built-in claude backend
[[backends]]
name = "claude"
env = { ANTHROPIC_BASE_URL = "https://proxy.example.com" }

# Custom name aliasing claude
[[backends]]
name = "bailer"
kind = "claude"
env = { ANTHROPIC_BASE_URL = "https://bailer.example.com" }

# Wrap the claude binary through a process wrapper
[[backends]]
name = "claude-wrapped"
kind = "claude"
cmd = ["wrapper", "claude"]

templated cmd and env

cmd[] and env values support {{ placeholder }} substitution, rendered at bind time. A school can pre-roll a backend that points at a launcher script under its own clone without knowing the absolute path.

# School-relative launcher script
[[backends]]
name = "codex-ace"
kind = "codex"
cmd  = ["{{ school_dir }}/skills/ace-connect/scripts/codex.sh"]

# Path-valued env var
[[backends]]
name = "claude-scoped"
kind = "claude"
env  = { XDG_CONFIG_HOME = "{{ project_dir }}/.config" }

Selectable via backend = "bailer" in any config layer or -b bailer on the CLI. Layer order for merging [[backends]] decls is school → user → project → local; later layers may add new entries or partially override earlier ones. A custom backend cannot introduce new behavior beyond its aliased kind — that requires a source change.

ACE links school folders into the project's backend directory. skills/ becomes a real directory with per-skill symlinks (driven by skills/include_skills/exclude_skills in ace.toml); the rest are whole-directory symlinks. Not all backends support every folder.

Trust modes change how the backend handles permission prompts: default uses the backend's standard handling, auto lets the model decide what needs approval, and yolo disables prompts entirely. Three equivalent ways to set the mode:

scope

Without a scope flag, trust writes to ace.local.toml (per-machine, gitignored — safe for personal preference). Add a global flag to target a different layer.

# Per-machine yolo for the current project (default scope)
$ ace yolo

# Project-wide auto — every clone of this repo picks this up
$ ace auto --project

# User-wide auto across every project on this machine
$ ace config set trust auto --user

# Inspect which layer is winning and why
$ ace config explain trust

per-backend mapping

ACE registers [[mcp]] entries from school.toml into the active backend. All entries are remote HTTP endpoints.

ACE verifies the backend is authenticated before launching a session.