// Persistent Redis-backed workspaces for agents. Use via `afs mcp`, the `afs` CLI, sync mode, live mounts, and explicit checkpoints.
Use when agents need persistent shared storage, when saving or restoring workspace state, or when coordinating file access across multiple agents and machines. Creates Redis-backed workspaces, checkpoints and restores agent state, mounts shared filesystems locally, searches workspace contents, and forks workspaces for parallel work.
Use when the user wants to migrate Codex state in ~/.codex into Agent Filesystem and mount the same shared Codex memory/settings across multiple computers. Recommends a .afsignore before migration and defaults to excluding worktrees, caches, logs, and temporary files.
"Use before writing, modifying, or reviewing code when the {{serverName}} MCP server is connected. Read the relevant standards files live from standards/, cite the specific path you used, and treat the workspace as read-only."
"Use when Codex has access to the {{serverName}} AFS shared-memory MCP server, or when the user asks to check shared memory, prior cross-agent context, durable learnings, memory entries, workspace memory, saved commands, acronyms, aliases, shorthand such as CPP/CCP, /memory-search, or /memory-record. Before any non-trivial answer in a connected project, fresh-search /shared-memory for the task's key nouns, read relevant entries, and build on them instead of re-deriving. After discovering a durable reusable fact, write a new dated memory entry and append a pointer to /shared-memory/index.md. Skip trivial, purely conversational, or session-only tasks."
"Use when answering non-trivial questions in a project connected to the {{serverName}} MCP server (tools prefixed {{toolPrefix}}*). Read wiki/index.md first, search wiki/ for key nouns, cite relevant wiki paths, and offer to file durable answers back into the Shared LLM Wiki (Karpathy style). When ingesting sources from raw/, update source/topic/entity pages plus wiki/index.md and append to wiki/log.md."
"Use when coordinating project work through the {{serverName}} MCP server: reading the spec, claiming tasks, updating in-progress work, recording done work, or writing open questions. Preserve per-owner task files and append-only shared docs."
| name | agent-filesystem |
| description | Persistent Redis-backed workspaces for agents. Use via `afs mcp`, the `afs` CLI, sync mode, live mounts, and explicit checkpoints. |
AFS is a workspace system for agents, backed by Redis. Use it when you want a durable workspace that still feels like normal files and directories, with explicit checkpoints and easy movement between MCP, sync mode, and live mounts.
Use for:
afs fs grep or MCP file tools are usefulAvoid for:
redis-cli examples from module-era docsafs mcpUse afs mcp when the agent can talk over MCP and does not need a local
directory.
afs CLIUse sync mode when the agent or user wants a real local directory:
./afs ws mount my-project ~/my-project
cd ~/my-project
Use ./afs config set --mode mount before mounting when you need the
workspace exposed directly as a mount rather than through the sync daemon.
./afs ws create my-project
./afs ws import my-project ./existing-dir
./afs ws mount my-project ~/my-project
cd ~/my-project
./afs fs grep --workspace my-project "TODO auth"
./afs fs grep --workspace my-project --path /src -E "timeout|retry"
./afs cp create my-project before-refactor
./afs cp list my-project
./afs cp restore my-project before-refactor
./afs ws fork my-project my-project-experiment
afs mcp and the CLI operate on the same workspace model..afsignore controls what gets imported from an existing local directory.