Skip to main content

Claude Code

Claude Code should be treated as a remote HTTP-first client for VibeRecall. That keeps the setup aligned with the broadest supported hosted path and avoids depending on capability surfaces that may vary between Claude integrations.

Use:

  • one remote HTTP MCP server
  • one project-scoped bearer token
  • one small tool subset for normal work

Do not design the core workflow around prompts or resources, even if a particular Claude integration happens to expose them.

Representative setup​

A representative command shape looks like this:

claude mcp add --transport http viberecall https://api.<your-domain>/p/<project_id>/mcp \
  --header "Authorization: Bearer $VIBERECALL_TOKEN"

If you also maintain a local helper:

claude mcp add viberecall-bridge -- vr-bridge serve-mcp

Treat the bridge as optional. The hosted memory service should remain usable without it.

Start with:

  • viberecall_get_status
  • viberecall_get_context_pack
  • viberecall_search_memory
  • viberecall_get_fact
  • viberecall_get_facts
  • viberecall_search_entities
  • viberecall_get_neighbors
  • viberecall_explain_fact
  • viberecall_save_episode
  • viberecall_get_index_status

Add only when the workflow needs them:

  • viberecall_update_fact
  • viberecall_index_repo
  • viberecall_working_memory_get
  • viberecall_working_memory_patch

Avoid broad daily-driver exposure for privileged maintenance tools such as entity merge or split.

Safe Claude Code workflow​

Start of task​

  1. Call viberecall_get_context_pack.
  2. If the task revolves around a known component, service, or entity, call viberecall_search_entities.
  3. Use viberecall_get_neighbors only once the task is anchored on the right entity.

During work​

  • save meaningful observations with viberecall_save_episode
  • use viberecall_search_memory for follow-up retrieval only when the task has shifted materially
  • avoid flooding the runtime with repeated broad searches

Before a code-heavy refactor​

  1. Check viberecall_get_index_status.
  2. If the latest index is stale or missing, trigger or request viberecall_index_repo.
  3. Retry the task context retrieval after the index is ready.

Prompts and resources policy​

Some Claude integrations can surface prompts, resources, or both. Public-safe guidance is still:

  • tools are the required baseline
  • prompts and resources are optional and compatibility-dependent
  • any valuable prompt workflow should still be achievable via direct tool calling

If tools work and prompts do not appear, the platform is still usable.

Common Claude Code mistakes​

  • Assuming a prompt-based workflow is required before the tool path is valid.
  • Treating missing resources as a server outage.
  • Reusing a stale session after a restart.
  • Over-enabling write or admin tools in a normal developer profile.
  • Assuming the hosted server can inspect local uncommitted files directly.

Local dirty workspace story​

The same rule applies here as with Codex:

  • a hosted MCP server cannot read your machine's local path directly
  • index from Git when possible
  • otherwise use a local helper or bundle flow and repo_source.type = "workspace_bundle"

See Local Workspace Bridge for the supported pattern.