Skip to content
v0.3.1Apache-2.0

Persistent Julia operator surface with state revival, workspace routing, and explicit host authority.

Palette plugin

This directory supplies the portable Agent Plugins manifest and a stdio MCP configuration. Install the runtime before enabling it. The template calls palette-mcp on PATH; the installer creates that command and can generate an absolute-path configuration for clients that do not expand plugin paths.

From the Palette repository root, after installing the runtime prerequisites:

julia --project=. -e 'using Pkg; Pkg.instantiate()'
cargo build --manifest-path runtime/host/Cargo.toml --release --locked
python3 runtime/security/prewarm_depot.py --project-dir "$PWD" \
  --host-bin "$PWD/runtime/host/target/release/palette-host"
python3 runtime/security/install_operator_plugin.py --format portable \
  --plugin-dir "$HOME/.local/share/palette-plugin" \
  --workspace-dir /absolute/path/to/workspace \
  --bin-dir "$HOME/.local/bin"

Cold preparation can take several minutes. Complete it before connecting an interactive client; see preparation and updates.

Import that generated directory in an Agent Plugins-compatible client. For the Codex compatibility format, omit --format portable and choose a separate plugin directory. Generic stdio MCP clients can run palette-mcp. First install and start the user service described in installation; the launcher connects to its stable runtime socket and does not own the Julia daemon lifecycle.

The repository marketplace is .agents/plugins/marketplace.json. Install the runtime and put the generated palette-mcp command on PATH before enabling its Palette entry. Marketplace availability depends on the client.

ChatGPT Chat requires a registered remote connection or Secure MCP Tunnel to this service; local files alone do not make a cloud conversation reach it. Public directory publication requires a supported hosted endpoint and review. No personal tunnel IDs, connection IDs, credentials, extra kernels, or external SDKs are distributed here.

The server offers palette, palette_control, palette_workspace, and palette_patch. These tools share the core runtime and authority contracts. A client lacking form confirmation cannot automatically approve exact patches.

Packaging follows the OpenAI plugin format. Connection setup follows OpenAI's connection guide.