Search
ProBM25-ranked FTS5 over every message in every session you have ever run. Sub-millisecond, highlighted snippets, #tag filters.
The memory layer for the agent you're steering.
You spend a quarter of your day steering Claude Code and Codex. Both forget all of it. Claude Recall is the substrate that captures every session from either agent, every tool call, every course correction, and pipes them back into the next conversation.
Install in two commands
npm install -g @clauderecallhq/clirecall start && recall openRequires Node 22 or newer. Ships as @clauderecallhq/cli on npm. Also available as a VS Code extension and an MCP server.
Free 7-day Pro trial
Index thousands of Claude Code and Codex sessions across every repo. For free.
Email only. No card. No auto-charge after the 7 days.
~/code
$
See it work
What every session looks like with memory.
One install covers Claude Code and Codex. Every session captured automatically. Searchable, graph-linked, available the next time the agent needs it.
Local-firstNo cloudNo silent telemetryNo signupWorks offline forever
The scale problem
30 terminals. 20 apps. Every session you open at 8am is gone by dinner.
VS Code's terminal scrollback caps around 1,000 lines. A real Claude Code or Codex session easily hits 10,000. When the scrollback evaporates, so does the reasoning, the debugging trail, the half-formed decisions.
Both agents already write every session to disk as JSON-L. Claude Code to ~/.claude/projects/, Codex to ~/.codex/sessions/. Either way there is no clean way to read, search, or reuse them. You end up copy-pasting terminal output into notes manually, or losing the context forever.
The bigger idea
You're not writing code. You're steering.
The hard part of an agent session was never the typing. It is the architecting, the discerning, the course-correcting: you set the goal, read what comes back, and steer it until it is right. That judgment is the one part of the work you cannot hand to the model, and it is the part neither Claude Code nor Codex carries into the next session.
Every session you've ever run contains that steering: your questions, your rebuttals, the moment you said "no, do it this way." Claude Recall is the substrate that captures all of it. Claude Code and Codex are both indexed from the first run, with no configuration for either, and the relevant pieces go back into the next conversation.
Over time the memory compounds. Your past steering teaches the next agent how you architect, whichever one you open, and the re-explaining that starts every session collapses into a single command.
The moat
One command replaces an hour of re-explanation.
Before Recall, starting a new session means re-pasting your stack, re-stating your constraints, re-building the context from scratch. After Recall, it's one pipe.
Without Recall
Open a new agent session.
Paste your stack.
Explain the auth flow again.
Re-describe the database schema.
Re-list the bugs you were debugging.
~3,500 input tokens. ~4 minutes.
With Recall
recall context abc12345 | claudeThe agent opens with the full context loaded.
That id can be a Claude Code session or a Codex one.
~700 input tokens. 1 command.
What it actually does.
Indexing every session and reading any transcript are free forever. The cards marked Pro need the one-time license.
Organize
Aliases, tags, markdown notes, pins, collections. Cross-project. Nothing is ever deleted. Three-layer durability with plain-text mirrors on disk.
Re-inject
ProPipe any past Claude Code or Codex session into a new conversation. CLI, web UI, or MCP tool call. This is the moat. Nothing else in this category has it.
Four surfaces, one memory.
Meet the product wherever you work: CLI, UI, model, or IDE.
One index sits behind all four, and it holds your Claude Code and Codex sessions side by side. Codex is indexed by default with no setup, badged by source, filterable in the CLI, the web UI and the MCP tools, and priced at OpenAI list rates.
CLI
Pipe-able subcommands you can script, plus a smart welcome dashboard and a full terminal UI when you want to browse by arrow key. recall list takes --source, so you can pull Claude Code sessions, Codex sessions, or both. Daemon-optional.
recall tui # arrow-key browse + live search
recall search "worker leak" # scripted query
recall list --source codex # claude-code | codex | allWeb UI
Three-pane dark-mode app served from the local daemon. Virtualized transcript renders 10k-message sessions instantly, and live rows sit above your history showing which editor terminals are still running an agent right now. A Source control narrows the list to Claude Code or Codex, and every Codex row carries a Codex chip.
recall openMCP server
Claude Desktop and Claude Code call search, get_session, context_for_session as native tools, plus locate_sessions to find the conversation you half-remember and bucket_search to read your own docs. search and list_sessions take a source argument, so one tool call can scope to Claude Code, to Codex, or to everything you have ever run.
{"mcpServers":{"recall":{"command":"claude-recall-mcp"}}}VS Code · Cursor · Windsurf
The editor hands over the tab name the moment a session starts in a terminal, so the link is recorded rather than guessed from the process tree afterwards. Claude Code and Codex sessions both inherit it, Codex on macOS and Linux. One setting, not zero: claude-recall.autoAlias ships off until you turn it on. Indexing, titles, search and pricing never need the extension.
recall install-extensionWho this is for
If you're steering agents every day, this was built for you.
Vibe coders
You want the clickable UI, zero terminal jargon, and a one-click Copy-as-context button that just works.
Power users
You juggle 20+ projects and 30+ terminals daily. Recall is your scrollback buffer across all of them.
Agentic engineers
You build workflows that orchestrate agents. Recall is the memory substrate they were missing.
Context engineers
You do prompt surgery. Recall gives precise per-session context with --since, --prelude, and tag filters.
Research labs
You reproduce experiments. Every session archived with three-layer durability and plain-text mirrors.
Frontier labs
You ship at the edge. Your sessions contain irreplaceable debugging trails. Recall makes them searchable.
Craftsmanship
Details you can feel.
Three-layer durability
Every write lands in SQLite, a plain-text mirror on disk, and never touches the source JSON-L that Claude Code and Codex wrote. Delete our database file and your aliases, tags and notes are still on disk in plain JSON and markdown, readable without us. Putting them back is a step you run deliberately.
Deterministic correlation
The VS Code / Cursor / Windsurf extension matches terminal tabs to session UUIDs by walking the process tree. A Codex rollout file stays open while the thread runs, so on macOS and Linux the open descriptor names the writing session outright and no timing window is needed.
Dynamic ports
The daemon picks a free ephemeral port and actively avoids 3000, 5173, 8080, so it never collides with the dev servers you are already running.
Auto-redacted secrets
If you ever paste an API key into an agent terminal, the indexer scrubs it before anything touches SQLite, so your secrets never end up searchable. Both agents go through the same scrubber.
Local-first, forever
Indexed on your machine. Nothing leaves silently.
Your session content stays on your machine.
Nothing leaves silently. The daemon binds to 127.0.0.1 only, and indexing, search and context re-injection are fully local. No signup. No cloud sync. Every outbound call is user-initiated or opt-in and off by default, apart from the license activation and daily license-revocation check that come with Pro. The README network-activity table lists every one, including the optional install ping (off by default).
Works offline on a plane, in a SCIF, on a locked-down laptop. Anywhere a developer ends up actually working.
Dead-man clause: a future version of Claude Recall can never delete your past. Your Pro license is validated once at activation, then works offline forever. If clauderecall.com ever disappears, every existing license stays valid.
Stop paying interest on your context debt.
One command, two minutes, and every Claude Code and Codex session you have ever run becomes searchable. You already pay for the agent. Add the memory layer once, less than two coffees a year out.
Get started
npm install -g @clauderecallhq/clirecall start && recall openRequires Node 22 or newer. Ships as @clauderecallhq/cli on npm. Also available as a VS Code extension and an MCP server.
A future version can never delete your past.