Claude Grep
coolcorexix/claude-grep
Busque e retome conversas do Claude Code por conteúdo em millisegundos
Ver no GitHubAgent CLI que busca o histórico completo de conversas do Claude Code por conteúdo (não apenas títulos), usando grep + fzf para encontrar e retomar qualquer sessão instantaneamente. Sem índices, apenas busca em tempo real nos arquivos de transcrição armazenados localmente.
Instalação
npx convez-ai coolcorexix/claude-grep1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Encontrar conversas anteriores por algo que foi discutido no meio da conversa
- Retomar sessões de trabalho sem lembrar do título inicial
- Filtrar histórico por modo somente-usuário para refinar buscas
Veredito

Ferramenta essencial para quem usa Claude Code intensivamente. Resolve um problema real — o histórico fragmentado por diretório — com uma solução elegante e performática via ripgrep + fzf, sem overhead de indexação.
README
A tradução para o português deste README ainda está na fila.
claude-grep
Search your Claude Code conversation history — find any past session by what was said, then jump straight back into it.
grep+fzffor everything Claude Code has ever told you.
claude-grep is a tiny CLI (run with the ccfind command) that searches the text of every Claude Code conversation stored under ~/.claude/projects/, and lets you resume any of them with a single keypress.
Demo

Why
claude --resumeonly matches a session's title (its first prompt) — not what was actually discussed inside it.- Your history is also fragmented per working directory, so each
pwdshows a different partial list and older work is hard to find and continue.
claude-grep searches the full transcript body across every project, so you can find a conversation by something that came up halfway through it — and drop right back into it.
Install
Homebrew (recommended)
brew install coolcorexix/tap/claude-grep
Then run ccfind. (You'll also need Anthropic's claude CLI on your PATH.)
Nix
nix run github:coolcorexix/claude-grep
Or nix profile install github:coolcorexix/claude-grep, or add it to your flake. (fzf, ripgrep, and Python are bundled in.)
Manual
Requirements: fzf ≥ 0.38, ripgrep, Python 3.8+, and the claude CLI on your PATH.
# macOS dependencies
brew install fzf ripgrep
git clone https://github.com/coolcorexix/claude-grep
ln -sf "$PWD/claude-grep/ccfind" ~/.local/bin/ccfind # ensure ~/.local/bin is on PATH
Then run ccfind. (Or just copy the ccfind script anywhere on your PATH.)
Usage
ccfind
- Type a phrase — results refresh live as you type.
- ↑/↓ move · Enter resumes the selected conversation · Esc quits.
- Ctrl-U toggles user-only mode (prompt flips to
user>) — match only what you typed, ignoring AI responses and tool output. Great for narrowing results when you remember your own words but the phrase appears in dozens of AI replies. Start in this mode withccfind --user-only(or-u). - Each row shows the matched snippet (phrase highlighted), the conversation's working directory (dimmed), and the message timestamp (right-aligned). The preview pane shows the session id, directory, and time.
How it works
- Fast:
ripgrepnarrows thousands of transcripts to candidates in milliseconds; only those files are parsed in Python for clean, highlighted snippets. No index, nothing running in the background. - Searches real content: your prompts, Claude's replies, and tool results — matches inside a tool call (e.g. a shell command) still surface via a raw fallback.
- User-only mode (
Ctrl-U/--user-only): restricts matching to text the user actually typed — assistant replies, tool output, and harness-injected blocks (system reminders,!-command output) are excluded across all three sources. - Sub-agent aware: sub-agent transcripts (
<project>/<PARENT-UUID>/subagents/agent-*.jsonl) can't be resumed directly, soclaude-grepresolves them to their parent session and resumes that — using the parent's working directory, so it works even if the sub-agent ran in a now-deleted git worktree. - Deduped: one row per resumable conversation, newest first.
Multi-agent: Claude Code · OpenCode · Hermes
If you also use OpenCode or Hermes, ccfind searches their conversation history alongside Claude Code's — from the same UI. It auto-detects which agents are installed and tags each row by source:
[c]Claude Code — JSONL transcripts under~/.claude/projects/[h]Hermes — sessions under~/.hermes/sessions/and~/.hermes/profiles/<name>/sessions/[o]OpenCode — the SQLite db at~/.local/share/opencode/opencode.db
Press Enter on any row and ccfind routes to the right agent's resume command (claude --resume / hermes --resume / opencode --session). Sub-agent sessions (Claude or OpenCode) auto-resolve to their resumable parent.
ccfind --sources # see which agents are detected
ccfind --source claude # search only one
ccfind --source claude,hermes # or a subset
ccfind --user-only # match only what YOU typed (alias -u; Ctrl-U in the UI)
Port a session into Claude Code (Ctrl-X)
Press Ctrl-X on any [o] OpenCode or [h] Hermes row to convert that session into a fresh, resumable Claude Code transcript and jump straight into it. Under the hood ccfind reads the source (OpenCode SQLite or Hermes JSONL/JSON) through the canonical conversation IR (see below), writes a new JSONL under ~/.claude/projects/<slug>/<new-sid>.jsonl with forkedFrom provenance pointing at the source, then execs claude --resume.
Tool calls, reasoning/thinking, text content, and inline images all carry over. For OpenCode, per-message provider details (model, cost, tokens) are preserved in metadata. For Hermes, OpenAI-style tool_calls are converted to Anthropic-style tool_use + tool_result blocks per Claude's convention.
CWD selection: OpenCode sessions carry their own working directory and resume there. Hermes sessions don't, so the new Claude session opens in ccfind's current directory (where you launched the search).
Directions today: OpenCode → Claude, Hermes → Claude. Claude → OpenCode and Claude → Hermes are on the roadmap (both targets track per-session economics fields or runtime-specific metadata that Claude transcripts don't carry, so the writers need to synthesize them).
Branch any session at any message (prompt tree)
Like git checkout from a commit — but for AI conversations. In the search UI, hit Ctrl-B instead of Enter on any row to:
- open a picker of all user messages in that session,
- pick one as your checkpoint,
ccfindcreates a new Claude Code session with the conversation rewound to that point — and resumes you into it.
The forked session is a first-class Claude session (works with claude --resume, appears in your history) and is tagged with the same forkedFrom: {sessionId, messageUuid} metadata that Claude's own /btw command writes — fully compatible with Claude's native fork format.
Useful when:
- you went down a wrong rabbit hole and want to try a different direction from a specific message,
- a session covered multiple tickets and you want each one as its own clean thread,
- you want to keep the original session intact while exploring an alternative path.
MVP supports Claude Code only; OpenCode and Hermes branching coming next.
Move conversations between agents (experimental)
ccfind ships a small foundation — ccfind_core — for porting sessions across agent tools through a shared canonical IR. The same model is what cross-agent features (import/export, bidirectional fork, fan-out) will be built on.
ccfind-convert capabilities # see what each agent supports
ccfind-convert list opencode --limit 5 # pick a session
ccfind-convert convert --from opencode --to claude --session <opencode-sid>
# → emits a fresh resumable Claude transcript, prints:
# { "session_id": "...", "path": "...", "resume_cmd": ["claude", "--resume", "..."] }
Use --dry-run to see the canonical-IR summary without writing anything to ~/.claude/. Or dump to inspect the IR directly:
ccfind-convert dump --from claude --session <sid> --pretty
Status today: OpenCode → Claude Code works end-to-end. Claude → Claude works (useful for forking). Claude → OpenCode is deferred — OpenCode tracks per-session economics (model, provider, cost, token counters) that Claude transcripts don't carry, so the OpenCode writer needs to synthesize those. Hermes adapter is a stub.
No community spec exists for stored-conversation portability (MCP, A2A, ACP, and Agent Client Protocol all standardize live interaction). The closest message-shape standard is OpenTelemetry's GenAI semantic conventions — ccfind_core uses Anthropic-style block names internally (so the Claude adapter is near-identity), and an OTel-conformant view can be derived by renaming tool_use ↔ tool_call_request, tool_result ↔ tool_call_response, thinking ↔ reasoning.
FAQ
How do I search my Claude Code conversation history?
Run claude-grep (the ccfind command) and type any phrase. It searches across every transcript in ~/.claude/projects/ and lets you resume the matching session.
Can I search past Claude Code sessions by content instead of just the title?
Yes — that's the whole point. claude --resume only matches the session title / first prompt; claude-grep searches the entire transcript body.
How do I find and reopen an old Claude Code conversation?
Search for something you remember discussing, select the result, and press Enter. claude-grep cds into that project and runs claude --resume on the correct session.
Where does Claude Code store its conversations?
As JSONL transcripts under ~/.claude/projects/<project>/<session-id>.jsonl (sub-agent sessions live in a subagents/ subfolder). claude-grep reads these directly — no export or setup needed.
Does it work with sub-agent sessions?
Yes. Sub-agent transcripts aren't resumable on their own, so claude-grep automatically resolves them to the resumable parent session.
claude-grep vs. claude --resume
claude --resume | claude-grep | |
|---|---|---|
| Matches on | session title / first prompt | full transcript content |
| Scope | current project only | all projects at once |
| Find by something said mid-conversation | ✗ | ✓ |
| Resumes sub-agent sessions | ✗ | ✓ (via parent) |
| Live, as-you-type search | ✗ | ✓ |
Limitations
- The right-aligned timestamp is sized to the terminal width at search time; resizing the terminal corrects on the next keystroke.
- Results are capped at the 80 most-recent matches per search to stay responsive.
License
MIT — see LICENSE.
More tools by the same author: nemothecollector.dev/products