kendreaditya/parlai

parlai

Local CLI wrapping the user's personal AI chat history across 8 providers. Lives at ~/workspace/parlai, installed on PATH as parlai.

Providers

chatgpt · claude · claude-code · codex-cli · codex-desktop · gemini · aistudio · perplexity

Commands

parlai status                              # which providers are authed
parlai list <provider> [-n N]              # recent conversations (live API or local files)
parlai get <provider> <id> [-f md|json]    # full conversation as Markdown
parlai search "<query>"                    # fan out across every authed provider
parlai search -p <provider> "<query>"      # one provider
parlai search "<query>" --content          # also fetch full conversation bodies (deduped)
parlai search "<query>" --json             # JSONL output for piping to LLMs/scripts
parlai open <provider> <id>                # open in browser
parlai login <provider>                    # interactive cookie paste (Chrome auto-detect fallback)
parlai --verbose <cmd>                     # print warnings to stderr

Stateless. No sync, no local cache, no DB. Every list and search hits the provider's web API live, or for local providers (claude-code, codex-*) reads their JSONL files directly. Trade-off: repeated queries re-hit the API each time — but no staleness, no cache invalidation, nothing to forget to sync.

Date filtering with --since / --until

Every list/search/sync command accepts --since and --until. Format options:

• ISO: 2026-04-19, 2026-04 (year-month), 2026-04-19T16:30:00
• Relative: 7d, 2w, 3mo, 1y, 24h, 90m
• Special: today, yesterday

parlai list claude --since 7d              # last week
parlai list claude --since 14d             # walks the API's pagination back 2 weeks
parlai search "india" --since 2026-04-01 --until 2026-04-30
parlai search "taxes" --since 2025-04-17 --until 2025-04-21   # a year ago ± 2 days

For remote queries, --since walks the provider's pagination back (newest-first) and stops once it crosses the boundary — effectively server-side filtering even though no provider's API exposes a date parameter. --until is best paired with --since (otherwise we still walk all of history).

Workflows

Find a past conversation by topic

parlai search "<topic>" --json | head -20

Returns one JSONL object per match: {provider, id, title, url, snippet, updated_at}. For local-only providers (claude-code, codex-*, gemini), the search auto-falls-back to the local FTS5 index.

Get the full conversation body

parlai get <provider> <id>                 # Markdown
parlai get <provider> <id> -f json         # structured JSON

Search and read in one shot

parlai search "<topic>" --content --json   # each line includes the messages array
parlai search "<topic>" -p claude -c       # human-readable, full convo bodies

--content deduplicates hits by conversation and fetches each one's full body. Not cached — fresh every call.

Output format

Default: fzf-style — [provider] id title then indented snippet (FTS5 highlights wrapped in bold yellow). For machine consumption (piping to other tools, feeding to an LLM), always pass --json for one JSON object per line.

Auth

Most web providers (ChatGPT, Claude, Gemini, Perplexity) read cookies live from Chrome via browser_cookie3 — no setup needed if the user is logged in. AI Studio uses Drive OAuth via the gog CLI. Local providers (claude-code, codex-*) need no auth.

If parlai status shows ✗ for a provider, run parlai login <provider> for an interactive cookie-paste fallback.

Notes

• No persistence. No DB, no cache file. Only ~/.parlai/credentials.json (manual cookie fallback) is ever written.
• Source repo: ~/workspace/parlai. To rebuild after changes: cd ~/workspace/parlai && uv sync.
• Local providers (claude-code, codex-cli, codex-desktop) search by scanning their JSONL files directly (case-insensitive substring).