Skip to content

Claude Code with Lore

Claude Code is Anthropic’s terminal-based AI coding agent. Lore’s gateway is the recommended way to run Claude Code with persistent memory and gradient context management.

Install

Section titled “Install”

The hosted install script sets up Lore and Claude Code in one step:

Terminal window
curl -fsSL https://withlore.ai/install | bash

Then launch Claude Code through the gateway:

Terminal window
lore run

lore run starts the Lore gateway (if not already running) and launches Claude Code with the right env vars to route through it. The gateway auto-detects Claude Code on your PATH.

Manual setup

Section titled “Manual setup”

If you’d rather start the gateway yourself and launch Claude Code directly:

Terminal window
# Terminal 1
lore start


# Terminal 2
ANTHROPIC_BASE_URL=http://127.0.0.1:3207 \
DISABLE_AUTO_COMPACT=1 \
claude

DISABLE_AUTO_COMPACT=1 is mandatory. Without it, Claude Code’s built-in auto-compaction destroys the very context Lore is trying to preserve. The gateway already sets this when you launch via lore run.

What you get

Section titled “What you get”

Every Claude Code conversation is captured in Lore’s three-tier memory. Distillations run in the background, the recall tool is available, and your project knowledge is exported to .lore.md and AGENTS.md automatically. See the architecture overview for the full picture.

Per-harness notes

Section titled “Per-harness notes”
  • Project identity is injected automatically. When launched via lore run, the gateway injects X-Lore-Project (your cwd) and X-Lore-Git-Remote (your git remote, if any) as ANTHROPIC_CUSTOM_HEADERS. This lets the gateway attribute sessions to the right project even when Claude Code’s system prompt doesn’t include the path explicitly.
  • Custom upstream headers are passed through Claude Code’s ANTHROPIC_CUSTOM_HEADERS env var. Set it before launching Claude Code and the headers reach the upstream unchanged. This is the official Claude Code way to send custom headers to the upstream.
  • Beta-gated fields are preserved. The gateway forwards anthropic-beta from the original request so features like context_management work transparently.

Switching to a different harness

Section titled “Switching to a different harness”

Your project knowledge (.lore.md, AGENTS.md, the SQLite database at ~/.local/share/lore/lore.db) is shared across every supported harness. Switching from Claude Code to Pi or OpenCode on the same project preserves curated knowledge, distillations, and the AGENTS.md sync.

Desktop app

Section titled “Desktop app”

The Claude Code Desktop app is an Electron app (macOS + Windows) that wraps the same claude CLI in a GUI. For local Code sessions, the Desktop spawns the claude CLI as a child process and reads ANTHROPIC_BASE_URL from ~/.claude/settings.json — the same file the CLI uses.

Setup:

Terminal window
lore setup claude-code-desktop

This writes ~/.claude/settings.json with env.ANTHROPIC_BASE_URL and env.DISABLE_AUTO_COMPACT — the documented path the Desktop’s Code tab honors for Claude sessions. On Windows it also runs setx ANTHROPIC_BASE_URL … so dev/preview servers inherit the gateway URL too. Restart the Desktop app afterward.

Fallback — in-app Local env editor. If your Desktop build does not pick up settings.json (upstream bug anthropics/claude-code#67619), or you want dev/preview servers covered on macOS, set the values in the Desktop’s in-app Local env editor: open the environment dropdown in the prompt box → hover Local → click the gear icon → add (one per line):

ANTHROPIC_BASE_URL=http://127.0.0.1:3207
DISABLE_AUTO_COMPACT=1

Restart the Desktop app. These values are stored encrypted (Electron safeStorage), so Lore cannot write them for you — this step is manual.

Linux note: The Desktop app is not available on Linux. Use the claude CLI through lore run (see the Install / Manual setup sections above).

Verifying the route: Start lore start in one terminal, then open the Desktop. The gateway dashboard at http://127.0.0.1:3207/ui will show incoming requests from the Desktop. If the dashboard stays empty, settings.json wasn’t picked up — apply the in-app env editor fallback above and restart the Desktop.

Next steps

Section titled “Next steps”
  • Architecture — how temporal storage, distillation, and the gradient context manager fit together.
  • Configuration — full reference for .lore.json.
  • Custom upstreams — corporate proxies, LiteLLM, Cloudflare AI Gateway.