Free & open source · MIT license · 100% local

Your AI conversations are an archive.
Start mining it.

Mine your AI conversations: LLM Toolkit turns your Claude Code and Codex transcripts into a searchable, browsable, editable knowledge base — search it, revise it with AI, transfer it between harnesses, compact it on your terms, then extract agents, skills, and fine-tuning datasets. Everything stays on your machine.

  • SQLite + local embeddings
  • No cloud, no telemetry
  • macOS · Linux · CLI
  • v0.0.1 — early days, moving fast

Capabilities

One local database, four surfaces: a terminal CLI, a full web console, a REST API, and a native macOS app.

Search everything

SQLite FTS5 keyword search plus semantic vector search over local MiniLM embeddings. Find the thread where you solved auth middleware two months ago — in seconds.

llm-toolkit search "auth middleware"

Read like a document

Rendered markdown, syntax-highlighted code, Mermaid diagrams, LaTeX, collapsible tool calls and thinking blocks.

Edit without fear

Collapse, remove, reorder, or inject messages. Every edit creates a new version — your source JSONL is never touched.

Switch harnesses

Move and transfer conversations between providers — Claude ↔ Codex — with clone, rehome, and archive operations. Your history isn't locked to one tool.

AI-assisted revise

Rewrite, tighten, or expand any message with the provider of your choice — Anthropic, OpenAI, Groq, DeepSeek, and more. Your thread, better words.

Compact, on your terms

Simplify long conversations with user-controlled compaction. Far more powerful than the harness's built-in /compact — you decide what survives the cut.

Build fine-tune datasets

Tag message ranges with gold/silver/bronze labels and export to OpenAI, Anthropic, or raw JSONL formats.

One control plane

Manage skills, commands, and MCPs for every harness from one place — enable, disable, and target per project or globally.

llm-toolkit skill list
context archaeology digging through your archives

Install

Built from source today — a package-manager release is on the roadmap. Requirements: Node ≥ 18 and pnpm ≥ 8; the macOS app additionally needs macOS 14+.

1 Clone & bootstrap

git clone https://github.com/the-robot-lives/llm-toolkit.git
cd llm-toolkit
make install

Builds the embedded skill manager (Rust), installs the llm-toolkit launcher to ~/.local/bin, and sets up shell completions.

2 macOS app optional

make install-osx

Builds and installs LLM Toolkit.app to /Applications — a native SwiftUI host with a built-in server supervisor. (make macos-run to try it without installing.)

3 Or run from source

pnpm install
pnpm dev:api   # REST API  → localhost:3100
pnpm dev:web   # Web console → localhost:5173

Handy for contributing. The API serves the built web console on port 3100 in production mode.

packing your bags conversations travel light

Quickstart

First launch auto-indexes ~/.claude/projects/ and ~/.codex/sessions. Data lives in ~/.llm-toolkit.

$ llm-toolkit                # launch the web console
$ llm-toolkit recent 2h      # fast lookup, no server needed
$ llm-toolkit search "auth middleware"
$ llm-toolkit list           # browse indexed sessions
$ llm-toolkit skill list     # manage provider skills