Skip to main content

TUI

Editing policy, roles, agents and MCP integrations interactively is what caesar config is for — an OpenTUI-and-React interface with a single hard requirement: Bun, not Node. OpenTUI's rendering goes through Bun's FFI, and Node 24 has no equivalent to offer it.

Requires Bun

Without bun on the PATH, caesar config explains the situation and points to the equivalent subcommands rather than failing dryly:

$ caesar config
The configuration TUI requires Bun: OpenTUI renders through its FFI, which Node 24 does not allow […]. "bun" was not found in the PATH.
Install Bun (https://bun.sh), or use the equivalent subcommands:
  - caesar policy show   Effective policy (allow/deny, provenance).
  - caesar role list     Roles, fallback agents, the agent picked today.
  - caesar agents list   Agent catalog: presence, capabilities, authorization.

None of these entry points — the TUI, the subcommands mentioned above, or the MCP server — reads or writes configuration on its own account. They are all facades over the same @caesar/core, which alone merges the three layers described in Configuration and stays the single source of truth throughout.

  • Tab / Shift-Tab, or 1-4 — switch screens.
  • s — save the pending changes on the active layer. Nothing is written to disk before it, except the system prompt editor, which writes on Ctrl+S.
  • p — cycle the editing scope: global → project → local. Switching scope with pending changes asks for confirmation before discarding them.
  • ? — show the help overlay.
  • q or Ctrl+C — quit, with confirmation if changes are pending.

The active scope stays visible at all times, together with the file path that s will write — knowing "project layer" does not say where that is, and that is precisely what one wants to check before saving.

Agents​

A catalog table (presence, permission, capabilities) with the selected agent's detail below it: presence, version, capabilities, permission, the roles that employ it, and — for a declared agent ([[agent]]) — its editable fields (display name, binary, arguments, network arguments, directory mode, native read-only).

  • ↑↓ — pick an agent.
  • Space — allow / deny (writes the policy list; denied always wins).
  • m — the agent's default model, native agents included (the [models] table).
  • n — declare a CLI the native catalog does not know: the identifier is typed inline, then bin/args and the rest are refined in the detail panel.
  • x — remove a declaration.
  • Enter — edit a declared agent's fields.

Allowing and declaring are two different gestures: the first writes a policy list, the second adds an agent to the catalog — conflating them is the easiest misreading of this screen.

Roles​

The list of roles on the left, editing of the selected role on the right: name, purpose, agents in their fallback order, mode, isolation, network, model, timeout, and the system prompt itself — content included, not just its path.

  • ↑↓ — pick a role; n creates one, x deletes it (renaming and deleting stay reserved for a role the active layer declares itself).
  • On a field, Enter edits it, or cycles it for mode/isolation/network.
  • On "Agents": Enter opens the fallback order; Shift+J/Shift+K moves the selected agent, a adds one, r removes one. The agent caesar_delegate would pick today is marked "← picked".
  • On "Model": the model requested from whichever agent gets picked — overrides the [models] default, loses to an explicit --model/model:. Empty means the agent's own default.
  • On "System prompt": Enter opens the full-screen prompt editor (below); f changes the declared file path.

Policy​

The fields of [policy] in plain language rather than raw TOML keys — each carries a label, and the TOML key it corresponds to is shown once selected, for whoever also edits the file by hand. A value inherited from a less specific layer is marked (← global).

  • ↑↓ — pick a setting; Enter edits it, or opens the allowed/denied list.
  • Inside a list: a adds an agent, r removes one, Esc returns to the settings.

The screen keeps a permanent reminder on screen: denied always wins over allowed — an agent present in both is denied.

Integrations​

For each of the five MCP clients (claude, codex, copilot, opencode, antigravity): its registration status of the caesar server (registered / not registered / not verifiable — claude and codex have no reliable, side-effect-free status read) and, in the detail panel, a preview of what Enter will actually do before it runs — the command that would execute, or the file that would be merged and under which key, with the rest of that file preserved. It is the only screen that writes outside the project, at the client's own configuration; the preview exists so that is never a surprise.

  • ↑↓ — pick a client; Enter installs or updates the registration.

Prompt editor​

A full-screen editor for a role's system prompt — the text placed at the head of the context passed to the agent, before the task's objective. Opened from the Roles screen ("System prompt", Enter).

Two things set it apart from the rest of the TUI:

  • It writes immediately. Ctrl+S saves the file directly, outside the global s/scope mechanism — a prompt is a single file, not a three-layer setting.
  • It names the file the engine will actually read, its absolute path shown at the top — a role coming from the global layer resolves its prompt in the current project, never a file shared between projects.

Esc abandons the edit, asking for confirmation first if the text changed.

Next steps​

  • Configuration — the full [policy], [[role]], [[agent]], [models] reference this TUI edits.
  • Using from Claude Code — registering the MCP server, one of the things Integrations does.