Skip to main content

Codex

Codex works well with VibeRecall when you keep the setup boring:

  • one hosted MCP server over HTTP
  • one bearer token from the target project
  • one small daily-driver tool subset
  • one optional local bridge only if you need local dirty-worktree support

What Codex is good at here​

Codex is a strong fit when you want:

  • a hosted memory service shared across environments
  • project-scoped context retrieval over HTTP
  • optional local stdio helpers for workspace-specific flows
  • explicit tool allowlists rather than a giant default surface

Required​

  • viberecall-core over remote HTTP

Optional​

  • viberecall-bridge over local stdio for workspace bundle or local helper flows

The hosted core remains the source of truth for project memory. The bridge is only a helper for local workspace access patterns.

Representative Codex config​

Configuration details vary slightly by Codex release, but the shape should look like this:

[mcp_servers.viberecall]
url = "https://api.<your-domain>/p/<project_id>/mcp"
bearer_token_env_var = "VIBERECALL_TOKEN"
tool_timeout_sec = 45

[mcp_servers.viberecall_bridge]
command = "vr-bridge"
args = ["serve-mcp"]
enabled = true

If you do not need local bundle or workspace-helper flows, leave the bridge disabled.

Enable these first:

  • 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

Enable only when you explicitly need them:

  • viberecall_update_fact
  • viberecall_index_repo
  • viberecall_working_memory_get
  • viberecall_working_memory_patch

Keep disabled by default in normal developer profiles:

  • viberecall_merge_entities
  • viberecall_split_entity

Safe workflow for everyday coding tasks​

Start of task​

  1. Call viberecall_get_context_pack with the task description.
  2. If the task is entity-centric, call viberecall_search_entities.
  3. Use viberecall_get_neighbors only after you know which entity matters.

During investigation​

  • save meaningful findings with viberecall_save_episode
  • avoid saving every intermediate thought
  • prefer one refined context query over repeated broad searches

Before a large refactor​

  1. Call viberecall_get_index_status.
  2. If code context is stale, trigger viberecall_index_repo from a trusted workflow.
  3. Wait for the index to become ready before assuming code context is accurate.

When correcting a stale belief​

  1. Call viberecall_explain_fact.
  2. Inspect lineage and supporting episodes.
  3. Only then call viberecall_update_fact if correction is actually warranted.

Common Codex mistakes​

  • Enabling the entire tool surface when only a handful of tools are needed.
  • Spamming search_memory instead of starting with get_context_pack.
  • Using graph tools before identifying the relevant entity.
  • Reusing a dead session after the backend restarts.
  • Assuming the hosted core can read the local repository directly.

Local dirty workspace story​

This is the most common design mistake in agent setups.

Hosted viberecall-core cannot read /Users/alice/repo or any other local path from your machine. If Codex is working on uncommitted local changes, you have two valid options:

  1. Push the branch and index from repo_source.type = "git".
  2. Use a local helper or bridge that packages the workspace and then index from repo_source.type = "workspace_bundle".

If you need that second path, read Local Workspace Bridge.

Prompts and resources​

If your Codex path or extension surfaces prompts or resources, treat them as optional convenience. The core integration should remain fully usable through tools only.