@clauderecallhq/cli
Advanced tools
Never lose a Claude Code session again. Local, fast, searchable memory over every session you've ever run.
Stop re-explaining your project to Claude every morning.
Claude Recall indexes every Claude Code session you've ever run, surfaces what matters,
and pipes the right context back in with one command.
Your session content stays on your machine. No cloud. No account. Opt-in telemetry, off by default.
Claude Code is incredible until the conversation ends. Then you hit three walls:
Context AmnesiaEvery new You built that context pyramid once. Now you're building it again. |
Cost BlindnessClaude Code tells you nothing about spend after a session ends. Which session burned through tokens? Which project is quietly eating half your Anthropic bill? You're guessing. |
Session SprawlClaude Code scatters conversations into Dozens of projects. Hundreds of sessions. Zero discoverability. |
npm install -g @clauderecallhq/cli
recall start
recall open
That's it. Recall indexes every past and future Claude Code session into a searchable local database, gives you a beautiful web UI to explore them, and lets you pipe any session back into a new conversation:
# The moat: past session -> new conversation
recall context abc12345 | claude
Claude picks up where you left off instead of asking you to explain it again.
Every Claude Code session you have ever run. Searchable. Threaded. Re-injectable.
Pipe any past session straight into a new Claude Code conversation as condensed markdown. Claude picks up where you left off instead of asking you to re-explain it. Nothing else in this category has it.
recall context <id> | claude
recall context <id> --prelude "continue this" | claude
recall context <id> --since 2h | claude
Group related sessions into a DAG you can see, rename, and walk. Drag-to-reparent in the graph view. Right-click rename. Bulk-generate coherent titles across an entire thread with one click.
recall thread list
recall thread create "auth-refactor"
recall thread add <thread> <session>
FTS5 BM25 ranking across every message, every session, in milliseconds. Highlighted snippets, inline #tag filters, project + date scoping.
recall search "zod schema"
recall search "auth #auth-fix" -p MyProject
Local 768-dimension embeddings via bge-base-en-v1.5 ONNX. Three-lane RRF fusion (BM25 + summary + vectors) finds sessions by meaning, not keywords. Your code never leaves your laptop.
The embedder model (~110MB) is opt-in: install on demand with recall semantic install. Keeps npm install -g quick and avoids forcing the download on users who never touch vector search.
One-click vectorization from the web UI. Each project header has a 🧠 Vectorize button. Click → pre-flight dialog shows scope (X of Y eligible sessions), three-✓ cost guarantees (free, on-device CPU only, no network), depth selector (Quick / Standard / Full / Custom chunks-per-session), and a depth-aware ETA computed server-side per cap level so Standard and Full give honest, different numbers. Counters are per-repository: the Queued, Throughput, and ETA tiles only count chunks from the project you opened the dialog from, with a separate small disclosure line surfacing how many chunks are queued from other projects ahead of yours (the worker drains FIFO globally). While running, the dialog flips to monitor mode with a live progress bar, ETA tick countdown, observed-throughput (drained-in-this-session) feeding the estimate, and visible warm-up (yellow) + wind-down (red) progress bars with timing context for the parts that have no deterministic ETA (embedder load is 5-60s by hardware; worker batch wind-down is 5-15s). Three explicit confirmation states close every run: green ✓ Vectorization complete — N chunks embedded in Xs banner when the queue drains naturally, green ✓ Already vectorized — N{project} is already done pre-flight callout (replacing the gray "Nothing to do" text) when there's nothing to vectorize, and the existing green ✓ Stopped — worker halted and queue cleared badge for Stop. Stop itself is a global kill switch: it clears every project's queue and halts the worker. The vector worker does not auto-resume on daemon restart by default; enable semantic.autoResumeWorker: true in ~/.recall/config.json if you want background draining without a click. Free tier doesn't see the button; endpoint is Pro-gated.
Anonymous install ping (opt-in, off by default). On first run we ask once whether you want to send a single anonymous message per month so we can count real installs vs bots. No PII, no fingerprint, no cross-month linkage. Full payload at clauderecall.com/telemetry. Toggle any time with recall telemetry on / off / view.
recall similar <session-id>
recall semantic install # one-time model download (~110MB, on-device)
recall semantic reindex # vectorize on local CPU (idempotent, resumable)
recall semantic verify-spawn # diagnostic — confirms claude CLI honors the no-persistence flag
Two lanes. Tier-2 (the commands above: install, reindex, verify-spawn) is free, local, and never sends anything anywhere. Tier-1 (recall semantic on / backfill / auto-extract) shells out to your local claude CLI to summarize sessions and costs plan tokens at ~30 sessions/min: opt-in for users who want LLM-generated summaries on top of vectors. Default-off.
Switching backends in place. Pro users running an older ONNX-based vector index who upgrade to the llama.cpp backend can run recall semantic migrate to re-embed the existing corpus under the new backend. The command is resumable (SIGINT-safe), atomic at switchover, and retains a 30-day rollback window via recall semantic rollback-migration.
Point Recall at a folder of your own documents and search them by meaning from any repo, including from inside a Claude Code session over MCP.
recall bucket create my-books
cp ~/Documents/notes/*.md ~/.recall/buckets/my-books/
recall bucket ingest my-books
recall bucket search my-books --query "how does retry backoff work"
Buckets are global. One created while working in project A is searchable from project B, so a reference library follows you everywhere instead of living in one repo.
Three ways Claude uses a bucket over MCP, cheapest first, so a large corpus never floods the context window:
| Tool | Cost | What it returns |
|---|---|---|
bucket_list | ~200 tokens | What buckets exist, how big, what topics |
bucket_search | a few thousand | The best-matching passages |
bucket_expand | negligible | The surrounding text, in original document order |
Measured on a 20-book, 2.86M-token corpus with ground truth taken from the
corpus rather than from our own search results: 75 to 90% of questions put a
correct passage in the top 5, in 74ms. The lower end is the hardest case,
finding one exact phrase among 6,565 passages; ordinary questions land at the
higher end. Everything runs on-device with the same bge-base-en-v1.5 model
used for session search, plus classic keyword search. Nothing is uploaded and
there is no per-query cost.
v1 reads .md, .markdown and .txt. Your source files are never modified,
and deleting a bucket keeps them unless you pass --delete-files. Embedding
only ever happens when you run recall bucket ingest - nothing indexes in the
background.
Per-session and per-project token + dollar totals. Daily sparkline. Top-10 heaviest sessions. Know exactly where your Anthropic bill is going.
recall stats <id>
recall stats --project Tools --days 7
Hover any day. Exact cost. Exact tokens.
Your 5-hour and weekly limits at the bottom of every Claude Code session, for every account you use, not just the one you are signed into right now.
● work 5h 31% ██░░░ 3h55m │ 7d 89% ████░ 18h25m │ live
○ side 5h ~0% ░░░░░ fresh │ 7d 77% ████░ 1d12h │ 1h55m ago
● is the account you are signed into; ○ is a stored snapshot, with its age. Colour is keyed on what you have actually spent, so it does not cry wolf at 30%.
Recall offers this once, the first time you run it in a terminal. Say no and it never asks again.
recall statusline on # or off, which restores exactly what you had
recall statusline list # every account and where its label came from
recall statusline name side personal # rename one
recall statusline ascii # for terminals that render block glyphs double-wide
Your 5-hour and weekly numbers come from the rate_limits field Claude Code already hands the status line: no network calls, no credential reads, no endpoint.
The optional Fable weekly column is the one exception, and it is off by default. Turn it on with recall statusline fable on and it does two things the rest of the status line never does: it reads your Claude Code OAuth credential from the macOS login keychain, and it calls api.anthropic.com/api/oauth/usage at most once every 15 minutes to fetch your Fable budget. It is macOS-only, it asks before enabling, and recall statusline fable off stops it. See Privacy & Security for the exact payload.
If you already have a status line, yours is kept and wrapped: statusLine is a single command slot, so Recall never claims it without giving it back.
Every session links to commits authored in its working directory during its time window. Reverse-map any commit back to the session that produced it.
recall correlate
recall blame <sha>
Let Claude label your sessions automatically. Two modes: MCP mode (zero-setup, runs through your existing Claude Code subscription at no extra cost) or BYOK (paste your own Anthropic API key). Off by default, fully opt-in.
Expose Recall to Claude Desktop, Claude Code, or any MCP client as native tools. 16 read tools always on, 22 write tools opt-in with rate limiting and append-only audit logging.
recall mcp
recall mcp --allow-writes
Claude searches your history, pulls context, and decides: no copy-paste.
Hand-curated hierarchical collections that cut across projects. Flat tags. Markdown notes per session. Aliases. Pins. Cross-session diff. Every write is reversible: three-layer durability with plain-text mirrors on disk.
Drag-to-organize. Collections cut across projects.
Dark-mode three-pane layout. Virtualized transcript (10,000-message sessions open instantly). Cmd+K command palette. Cross-session diff. First-60-seconds onboarding tour. Export to markdown, HTML, or print-ready PDF.
Or stay in the terminal. recall tui ships the same library as a full-screen terminal app: slash to filter, arrow keys to navigate, live preview.
Three rotating cards in the left pane surface forgotten sessions worth revisiting: a 60+-day-old session that overlaps with what you've been working on, your top-cost session this week, and the session that authored your current HEAD.
recall digest
VS Code, Cursor, Windsurf, VS Code Insiders. Tag terminal tabs from inside your editor; terminal names sync to Recall so sessions get meaningful aliases with zero CLI ceremony, live as you rename tabs.
recall install-extension
mcp_audit_events with status (ok / error / rate_limited). Pure local, no telemetry.recall context <id> | claude) the moment you open the web UI.Cmd+Click toggle, Shift+Click range, right-click to copy IDs.~/.claude/projects/ are strictly read-only.| Key | Action |
|---|---|
/ | Focus search |
? | Open Command Center |
Cmd+K / Ctrl+K | Command palette |
Esc | Close / clear / dismiss |
n / p | Next / previous search match |
c | New collection |
Cmd+Shift+A | Add session to collection |
Cmd+Enter | Save note |
People sometimes ask if this is "Obsidian for Claude Code." It is not. A general note-taking app cannot watch your filesystem for new sessions, index JSONLs into FTS5 + vector search, expose those sessions as MCP tools, or pipe a session back into Claude with one command. Claude Recall is built specifically for the loop you actually run.
npm install -g @clauderecallhq/cli
# Start the background daemon (file watcher + local HTTP server)
recall start
# Open the web UI
recall open
The daemon watches ~/.claude/projects/ for new and changed session files, indexes them into a local SQLite database at ~/.recall/db.sqlite, and serves the web UI on a random 127.0.0.1 port.
Two interactive surfaces, both work without the daemon (they read SQLite directly):
recall # smart welcome dashboard with status + slash command palette
recall tui # full terminal UI: arrow-key browse, live search, preview pane
recall (no args) prints a branded splash with version / daemon / sessions / license status and drops into a slash-aware prompt; type / to browse 60+ commands grouped by purpose (Setup, Browse, Pipe to Claude, Threads, Inference, Analytics, Daemon, Diagnostics, Integrations, Sharing, Feedback, Pro).
recall tui mounts a full-screen terminal app modeled on lazygit and k9s. Arrow keys to navigate, / to filter, enter to view the full transcript, o to open the selected session in the browser, q to quit.
recall projects # list projects with session counts
recall list # 30 most-recent sessions
recall list -p Tools # filter by project
recall show <id> # pretty-printed transcript
recall search "auth bug" # full-text search
# Pipe a past session into a new Claude Code conversation
recall context abc12345 | claude
# With a custom instruction
recall context abc12345 --prelude "continue where we left off" | claude
# Just the last 2 hours
recall context abc12345 --since 2h | claude
Claude resumes with full prior context. One command. No re-explaining.
Don't think of Claude Recall as a separate tool you context-switch to. It's Claude Code's memory layer: you reach it from inside any session through four surfaces, in increasing power:
1. Bash tool: always available, zero setup.
Inside any Claude Code conversation, just say "run recall search auth bug" (or any other recall command). The Bash tool executes it; output lands in the conversation. Works the moment you npm install -g @clauderecallhq/cli. No config.
2. Slash commands (skills): one-tap workflows.
When Claude Recall is installed as a Claude Code plugin, typing / in Claude Code surfaces ready-made workflows:
| Slash command | What it does |
|---|---|
/claude-recall:recall <query> | Search past sessions and pull relevant ones in as context |
/claude-recall:context <id> | Load one specific session's full transcript |
/claude-recall:neighborhood <id> | Bundle parents + children + citations + similar |
/claude-recall:tag <id> <tag> | Tag a session for later retrieval |
/claude-recall:alias <id> <name> | Rename a session |
/claude-recall:threads-sync | Capture currently active terminals into a thread |
/claude-recall:thread-scan | Auto-detect threads across historical sessions |
/claude-recall:setup | Verify installation and MCP connection |
3. MCP server: the model calls Recall directly.
The most powerful surface. With recall mcp registered as an MCP server in Claude Code, the model itself decides when to call Recall mid-conversation. Ask "have I solved this auth bug before?" and Claude calls search autonomously, gets results, and uses them. Ask "pull yesterday's context" and it calls context_for_session. No commands, no slash menu, pure intent.
4. Piping into a fresh claude: start tomorrow with yesterday's context.
The cross-session continuation pattern, run from your shell:
recall neighborhood abc12345 | claude # bundle related sessions, start fresh chat
recall context abc12345 | claude # one specific session, start fresh chat
Use case: you finish work Monday night, restart Tuesday morning with a clean Claude Code window, and want it to know what happened yesterday without re-explaining.
Point any MCP client at Recall and the model calls it as native tools:
{
"mcpServers": {
"recall": {
"command": "claude-recall-mcp"
}
}
}
Read tools (16, always available):
| Tool | Purpose |
|---|---|
list_projects | Every indexed project with session/message counts |
list_sessions | Recent sessions; filters: project, tag, date range |
list_tags | Tag cloud with counts, most popular first |
search | Full-text search with #tag support; BM25 + optional vector fusion |
find_similar_sessions | Semantic similarity via vector embeddings (Pro) |
semantic_status | Health snapshot: model, worker, queue depth (Pro) |
get_session | Full transcript (accepts 8+ char id prefix) |
context_for_session | Condensed markdown export, ready to inject |
recall_neighborhood | Bundle parents/children/citations/similar as one budget-bounded markdown blob |
doctor | Read-only health snapshot: db size, WAL, FTS5 fragmentation, integrity check |
thread_list | All threads, newest first; can exclude archived |
thread_get | Full thread detail with every session edge |
thread_for_session | Non-archived threads referencing a session |
bucket_list | Inventory of local document buckets: names, document/chunk counts, top headings |
bucket_search | Semantic search inside one bucket; best-matching passages with heading path |
bucket_expand | Chunks surrounding a search hit, in original document order |
Write tools (22, opt-in via --allow-writes):
| Tool | Purpose |
|---|---|
list_sessions_to_tag | Surface untagged sessions for review (auto-tag flow) |
apply_tags | Add tags to a session (merge-mode, never deletes) |
add_tag / remove_tag | Tag management; removals logged in append-only log |
set_alias | Human-friendly session name; previous alias archived |
append_note | Add markdown to session note; separated by --- |
create_collection | New collection; optional parent, icon, color |
add_session_to_collection | Idempotent collection membership |
remove_session_from_collection | Remove session; logged in append-only log |
thread_create | Create thread, optionally seed with origin session |
thread_add_session | Attach session; role=origin or child |
thread_set_parent | Change parent within thread; null clears parent |
thread_remove_session | Detach session from thread |
thread_rename | Change thread display name |
thread_close / thread_reopen | Mark thread as closed or reopen |
thread_archive | Soft-delete thread; hidden by default |
thread_merge | Move all edges from source to dest, delete source |
thread_split | Peel sessions into new thread |
sync_active_sessions | Capture every Claude Code terminal currently open in a repo into one thread, so parallel sessions stay organized |
generate_thread_titles | Generate coherent titles for sessions in thread DAG |
optimize | WAL checkpoint, FTS5 segment merge, planner stats refresh; optional VACUUM |
All writes are rate-limited (default 60/min), zod-validated, and audited to ~/.recall/audit/.
~/.claude/projects/**/*.jsonl <-- Claude Code session files (read-only to Recall)
|
v chokidar file watcher
~/.recall/db.sqlite <-- indexed database (FTS5 + sqlite-vec 768d)
|
+-- background embedding worker (idle-aware, transactional)
+-- git correlator (read-only 'git log' scoped to session cwd)
+-- semantic / auto-tag scanner (opt-in, off by default)
|
+-----+-----+
| |
CLI HTTP server (Hono, 127.0.0.1 only)
| |
recall ... Web UI (React + Tailwind SPA, Vite-bundled)
MCP server (stdio, 38 tools, opt-in writes)
| Path | What |
|---|---|
~/.recall/db.sqlite | Indexed database (projects, sessions, messages, FTS5, vectors, aliases, notes, tags, collections, threads, semantic summaries, commit links, usage rollups, audit logs) |
~/.recall/aliases.json | Plain-text mirror of every alias with full edit history |
~/.recall/notes/<session>.md | One markdown file per session note |
~/.recall/tags.json | Current tag state + append-only event log |
~/.recall/collections.json | Current tree state + append-only event log |
~/.recall/semantic/<id>.json | Per-session summary + keywords (semantic mode) |
~/.recall/license.jwt | Activated Pro license (offline-validatable, dead-man clause) |
Claude Recall is local-first by design, not as an afterthought.
127.0.0.1 only. Never 0.0.0.0, never a public interface.claude -p on your own account with the session text). A fourth path, the optional Fable status-line column, sends a credential but no session content.~/.recall/ and nowhere else.~/.claude/projects/ are strictly read-only to Recall.Everything Recall can send over the network, and exactly what triggers it:
| Trigger | Endpoint | Exact payload | What it gets | Automatic? |
|---|---|---|---|---|
recall activate <key> | clauderecall.com/api/license/activate | license_key, machine_fingerprint (SHA-256 of hostname + username + OS + arch), instance_name (your hostname plus a random suffix) | A signed JWT for offline validation | No - only when you activate |
recall trial <CODE> / activating a promo code | clauderecall.com/api/trial/cli-redeem | promo_code, email (the address you type at the prompt), machine_fingerprint, instance_name, referral_source | A trial license | No - only when you redeem |
| Pro license revocation check | clauderecall.com/api/license/check | license_key - the key itself, not a fingerprint | OK / revoked status | Yes - 30s after daemon start, then every 24h, once licensed |
recall semantic install (or the Vectorize dialog's install button) | registry.npmjs.org, then huggingface.co → us.aws.cdn.hf.co | Nothing about you | The bge-base-en-v1.5 model files. The four model files and the GGUF are SHA-256 pinned and deleted on mismatch. On linux/x64 only, npm additionally pulls CUDA binaries for onnxruntime-node from api.nuget.org / aiinfra.pkgs.visualstudio.com - that transitive download is not hash-verified by us | No - only when you opt in to vector search |
| Anonymous install ping | clauderecall.com/api/install-ping | event, version, platform, arch, month, nonce (rotating; no cross-month linkage, no fingerprint, no PII) | Nothing | Yes - once a month, only after recall telemetry on (off by default; auto-declines in CI and non-TTY) |
| Share a recap card | clauderecall.com/api/cards | kind, title, description, width, height, pngBase64 (the rendered image). A session card's title, and text drawn into the image, can derive from your first message | A public recap URL | No - only when you click Share |
| Social share buttons | The network you pick | The recap URL and caption | Nothing | No - only when you click |
| Feedback form | clauderecall.com/api/feedback | score, comment (your message, verbatim), surface, version, os, trigger_kind, license_jwt - so Pro feedback is attributable to your license | Nothing | No - only when you submit |
| AI tagging with your own Anthropic API key | api.anthropic.com/v1/messages | Your top-50 tag vocabulary across the whole database, the session's current_tags, the project folder name (not the full path), alias, git_branch, first_user_message (in full - not truncated), and message_sample (5 messages sampled across the session, first 400 chars of each). Note alias carries your VS Code tab name, which may contain client or project names | Suggested tags | Off by default; needs your key and an explicit enable. With autopilot also switched on, tagging then runs in the background and is kicked on daemon start |
| Fable weekly column in the status line | api.anthropic.com/api/oauth/usage | Authorization: Bearer <your Claude Code OAuth token, read from the macOS login keychain>. No session content is sent | Your Fable weekly usage | Off by default, macOS only. After recall statusline fable on, at most once every 15 minutes while a session renders |
| ✨ Generate title / synopsis | Your own claude CLI | The session text being summarized, via a locally spawned claude -p under your own account | The generated text | No - only when you click Generate |
Activate is required for Pro features. Telemetry is opt-in. On a fresh install with nothing activated, enabled, or clicked, Recall makes no network calls at all; every call above is unlocked only by the action named in its Trigger column. Indexing, search, and context re-injection stay fully local either way. Full telemetry disclosure: clauderecall.com/telemetry.
# Surfaces
recall # smart welcome dashboard + slash command palette (no args)
recall tui # interactive terminal UI for browsing and search
# Lifecycle
recall start # start daemon (watcher + local HTTP)
recall stop # stop daemon
recall open # open web UI (starts daemon if needed)
recall status # db + daemon health
# Indexing
recall index # scan for new/changed session files
recall index --force # reindex everything
# Explore
recall projects # list projects with session counts
recall list # 30 most-recent sessions
recall list -p Tools # filter by project substring
recall list -n 100 # higher limit
# Read
recall show <id> # pretty transcript, auto-paged
recall show <id> --raw # raw JSONL
recall show <id> --no-pager
# Search
recall search "zod schema"
recall search "auth" -p MyProject
recall search bug -n 50
# Context re-injection
recall context <id> # condensed markdown
recall context <id> | pbcopy # to clipboard (macOS)
recall context <id> | claude # into new Claude session
recall context <id> --full # full transcript
recall context <id> --since 2h # last 2 hours only
recall context <id> --prelude "continue this" # prepend instruction
recall context <id> --subagents # include subagent messages
# Stats
recall stats <id> # tokens + dollars for one session
recall stats --project Tools # per-project rollup
recall stats --days 7 # 7-day overview
recall stats --backfill # one-shot usage backfill
# Git correlation
recall correlate # link sessions to commits
recall blame <sha> # commit -> session reverse lookup
# Semantic / vector search (Pro)
# Embedder model is opt-in — `npm install` does NOT auto-download it.
recall semantic install # download on-device embedding model (~110MB)
recall semantic status # model + backfill progress + auto-extract state
recall semantic reindex # re-embed everything
recall semantic migrate # re-embed existing corpus under a different backend (Pro)
recall semantic rollback-migration --force # restore prior corpus within 30-day window
recall semantic prune-rollback --force # drop backup table early
recall extract-outputs --project <name> # extract on demand through Claude (Pro); manual-only — background spend is hard-blocked
recall similar <id> # cosine kNN over session chunks
# Threads
recall thread list # all threads, newest first
recall thread create <name> # new thread
recall thread add <t> <s> # attach session to thread
# Rediscovery
recall digest # today's "For you" picks
recall digest --json # machine-readable
# MCP server
recall mcp # stdio, read-only
recall mcp --allow-writes # opt-in write tools
# Pro license
recall activate <key> # one-time activation, offline forever
recall activate RECALL7DAY # CLI-direct 7-day trial — prompts for email, no browser
recall trial RECALL7DAY # same, alternative entry point
recall license # show license status
# IDE extension
recall install-extension # auto-install for VS Code / Cursor / Windsurf
# Maintenance
recall health # memory health scores per project
recall paste # archive clipboard content into Recall (opt-in)
recall audit-secrets # scan index for residual secrets
| Layer | Technology |
|---|---|
| CLI + daemon | TypeScript (strict), Node 22+, ESM |
| HTTP server | Hono + @hono/node-server |
| Database | node:sqlite (Node's built-in SQLite) with FTS5 full-text search |
| Vector store | sqlite-vec (768d, on-device) |
| Embeddings | @huggingface/transformers (bge-base-en-v1.5 ONNX) |
| File watcher | chokidar |
| CLI parsing | commander |
| MCP server | @modelcontextprotocol/sdk (stdio) |
| Validation | zod |
| Auth | jose (offline-validatable signed JWT licenses) |
| Web UI | Vite 8 + React 19 + Tailwind CSS |
| Data fetching | @tanstack/react-query |
| Virtualization | @tanstack/react-virtual |
| Markdown | marked + highlight.js + DOMPurify |
Claude Recall is a Node.js CLI with a handful of native dependencies: sqlite-vec (vector search) and sharp (share-card image rendering), plus the optional @huggingface/transformers for semantic search. Each ships as a family of per-architecture packages (for example sqlite-vec-darwin-arm64, @img/sharp-linux-x64) that npm resolves at install time to match the machine running the install. No compile toolchain required.
Because the binding is chosen at install time, not at runtime, the thing that breaks Recall is reusing an existing node_modules after switching Node's architecture, not switching Node itself. If you move a checkout between an Intel Node and an Apple Silicon Node (or between x64 and arm64 generally), reinstall dependencies with the new Node on PATH rather than carrying the old node_modules over. recall doctor detects the mismatch and points back here.
| Platform | Tier | Status |
|---|---|---|
| macOS arm64 (Apple Silicon, M1+) | Tier 1 | Tested in pre-publish smoke test |
| macOS x64 (Intel) | Tier 1 | Tested in pre-publish smoke test |
| Linux x64 (glibc, kernel 4.x+) | Tier 1 | Tested on every CI run (Node 22, 24) |
| Linux arm64 (glibc) | Tier 2 | Best-effort. Native deps ship arm64 prebuilds; report issues if you hit one |
| Windows x64 (Windows 10+) | Tier 1 | Tested in pre-publish smoke test |
| Windows arm64 / Linux musl (Alpine) / 32-bit | Unsupported | Native deps lack reliable prebuilds; install will be rejected |
Node: 22 LTS or 24 LTS. Node 20 and earlier are unsupported (declared in engines.node).
Semantic search is optional. The on-device embedder is the only feature that loads the native ONNX runtime. The model (~110MB) is opt-in: install on demand with recall semantic install. Core CLI features (search, list, context, daemon, MCP) work on every supported platform regardless of whether the embedder is installed. If the embedder fails to load on your platform, you get a clear error pointing here, and the rest of Claude Recall keeps working.
Claude Recall ships on every surface a developer already works in.
| Surface | Identity | Install |
|---|---|---|
| npm (CLI + MCP) | @clauderecallhq/cli | npm i -g @clauderecallhq/cli |
| VS Code Marketplace | clauderecallhq.clauderecall-vscode | Search "Claude Recall" in Extensions |
| MCP Registry | com.clauderecall/recall | Auto-discovered by MCP clients |
| Claude Code plugin | claude-recall | /plugin marketplace add …claude-recall-plugin.git then /plugin install claude-recall@claude-recall |
| Website | clauderecall.com |
The npm package ships the CLI and the MCP server. The VS Code extension is a thin companion that talks to the local daemon over 127.0.0.1. The MCP Registry and Claude Code plugin both point back to the npm package: one source of truth, one update pipeline.
To install the Claude Code plugin, add the marketplace then install (the .git
suffix is required: Claude Code git-clones the marketplace; a bare URL will not
resolve):
/plugin marketplace add https://gitlab.com/clauderecallhq/claude-recall-plugin.git
/plugin install claude-recall@claude-recall
One tier is free forever. One tier is a one-time purchase that works offline for life, even if we disappear. One tier is a conversation.
$0. Forever. No signup. No credit card.
~/.claude/projects/ JSON-Llist, show, projects, statusnpm install -g @clauderecallhq/cli
$49.69 one-time. Pay once. Offline forever. Lifetime updates. Dead-man clause included. Never a subscription.
recall context re-injection, the moatBuy Pro for $49.69 (one-time, lifetime)
Contact us. Built for 50+ developer organizations.
| Tool | Price |
|---|---|
| Claude Recall | $49.69 one-time |
| Claude Code | $20 / mo |
| Cursor | $20 / mo |
| Raycast Pro | $8 to $16 / mo |
| GitHub Copilot | $10 / mo |
Competitor prices subject to change.
Your Pro license validates against our server exactly once, at activation. After that, Recall works offline forever. If clauderecall.com ever disappears, every existing license stays valid; we'll publish the validation secret if we ever shut down, so you're never stranded.
Full pricing FAQ: clauderecall.com/pricing
The highlights below are the user-facing changes from recent versions. The CLI, MCP server, and daemon ship together in the npm package and are versioned independently; see the full changelog for every entry, including dependency and internal updates.
0.124 to 0.127 - Survives an architecture switch. Changing Node architecture under an existing node_modules used to crash every recall command at module load, including recall stop and recall doctor - the two commands you need to recover. Now you get a plain-English warning before each command, doctor reports a full Architecture section even when the database cannot open, and image commands explain the mismatch instead of surfacing a native library's own advice. Security: the RECALL_API_BASE development override is now pinned to loopback, so a hostile environment variable can no longer redirect license traffic off-machine, and the daemon's claude -p spawns no longer pass --permission-mode bypassPermissions - they run under your own permission configuration instead of overriding it. New optional Fable weekly column in the status line (off by default, macOS only; it reads your Claude Code credential from the keychain and calls Anthropic - see the network table above). Also: a crashing panel no longer blanks the whole web UI, and Vectorize can be cancelled promptly. Correction: the network-activity table above has been rewritten to name every payload field exactly; several earlier rows understated what they send, and the optional Fable column was missing entirely.
0.123 - Security hardening: the embedder runtime can now only be installed or repaired from the CLI (recall semantic install); the web UI can no longer trigger a runtime install. The Vectorize dialog still detects a missing or broken runtime, but now shows the exact command with a copy button instead of running the install itself. Vectorizing also got faster: a real embedder thread cap (about half your cores by design), no more dead two-second pauses between sessions, and Stop responds within seconds instead of after everything in flight.
0.122 - Usage caps in the Claude Code status line: your 5-hour and weekly limits at the bottom of every session, for every account you use, not just the one you are signed into. Built entirely from the rate_limits data Claude Code already hands the status line; as shipped in 0.122 this involved no network calls and no credential reads. (0.124 later added an optional Fable column that does both, off by default - see above.)
0.120 - Transcript toggles that actually hide: tool calls, tool results, and plumbing rows each cycle through shown, collapsed, and fully hidden, with a receipt bar naming exactly what is withheld ("447 hidden: 295 tool calls, 152 plumbing") and one click to restore everything. PDF export got a fidelity overhaul: it now prints what the transcript is actually showing, honors your toggles, and typesets tables, symbols, timestamps, and spacing properly instead of producing a wall of monospaced text.
0.115 to 0.121 - The Decisions tab answers a question the product could not answer before: how many prompts did you personally type? Pick Today, 7d, 30d, or All time, per repository or across all of them, and get the count, a per-repo table, a sparkline, and a median active-hours-per-day tile. Claude Code writes tool results, notifications, and other machinery into the same user role your typing lands in (a raw count overcounted by about 14x on one measured day), so the filtering is the whole feature.
0.103 - One-click PDF export: any session view can now produce a clean, conversation-only PDF. MCP search responses include a short tip when results look thin, nudging the agent to retry with more specific wording or a project scope, so multi-turn lookups converge on the right session faster. License signing rotated to a fresh key; existing licenses keep validating unchanged, no action needed.
0.102 - FleetView: a live per-repository roster of your active Claude Code sessions with context fullness at a glance, so you can see which terminal is doing what and how full each context window is.
0.101 - Search results got measurably better: the ranking fusion that merges keyword and semantic lanes was retuned, improving top-5 relevance by 11.6 points on our internal retrieval benchmark. Also a supply-chain hardening pass: cleared an adm-zip CVE, deduped a nested sharp, and added a cross-platform supply-chain gate that runs before every release.
0.97–0.100 - Buckets: point Recall at any folder of documents (recall bucket create / scan / search) and search it with the same hybrid keyword + semantic retrieval the session index uses; worst-case retrieval quality improved from 0.300 to 0.775 Hit@5 during development. The embedding model now stays warm in long-lived processes, so repeat semantic queries skip the model reload entirely.
0.96 - Semantic search works again once your history gets big. Past a certain size the vector index would refuse to run (corpus_too_large) and quietly fall back to keyword search. The index is now sharded by project, so a scoped search only scans that project instead of your entire history, turning searches that used to be refused or hang into ~13–106 ms. Run recall semantic migrate --partition to switch over: it's proven on a real 201,931-vector corpus (zero vectors lost, ~3-minute migration, full backup + parity check before anything moves, all local, no telemetry). recall semantic status now reports the daemon's real embedder/worker state instead of a phantom "not loaded." Vectorization is now strictly manual and button-only: live session activity never auto-enqueues embedding work, so the daemon no longer sits at "warming up" indefinitely or spends CPU in the background. A hard kill-switch also guarantees that no background claude -p process is ever spawned without an explicit, user-initiated action, so indexing and titling never quietly bill against your plan.
0.95.13 - Batch-spawned claude -p sessions from temp directories no longer flood the session/project lists with thousands of identical entries, and the daemon no longer exhausts file descriptors when those spawn dirs pile up. Spawn transcripts stay on disk; pass ?system=1 to view them.
Looking for a specific version or the complete history? The full changelog carries every release with
Added/Fixed/Internalsections, in Keep a Changelog format.
Proprietary. Free tier is free forever for personal use. Pro is a one-time purchase with offline-validatable JWT and a dead-man clause: if clauderecall.com ever shuts down, every existing license stays valid and the validation secret will be published so you're never stranded. Full terms at clauderecall.com/terms.
FAQs
Never lose a Claude Code session again. Local, fast, searchable memory over every session you've ever run.
The npm package @clauderecallhq/cli receives a total of 544 weekly downloads. As such, @clauderecallhq/cli popularity was classified as not popular.
We found that @clauderecallhq/cli demonstrated a healthy version release cadence and project activity because the last version was released less than a year ago. It has 1 open source maintainer collaborating on the project.
Security News
GPT-6 Astra hits 100% on ExploitBench and finds zero-days autonomously, while independent tests reveal scope violations and monitoring gaps.
Product
Socket can now send alerts and supply chain attack notifications to Microsoft Teams, with filters that route the right updates to each channel.
Security News
pnpm 12 rewrites the package manager in Rust, cutting install times by up to 90% while preserving pnpm 11 workflows and lockfiles.
