learnlance

learnlance [command] [options]

learnlance is a Python CLI that watches coding agents and turns the code they write into a persistent personal knowledge graph. After each turn that edits files, it extracts transferable concepts (for example delta encoding, not "this project's helper"), merges them into a per-project graph, and regenerates a self-contained HTML map.It integrates with Claude Code, OpenAI Codex, Cursor, GitHub Copilot (CLI, cloud, and VS Code), Command Code, Kiro, Gemini CLI, Google Antigravity, and a git post-commit fallback. Most hooks are written per project; the Claude Code hook is user-level. setup (and a bare install) detects which agents are present and configures them. Later commands also run a quiet auto-setup so moving to a new project does not require an extra step, except setup, install, uninstall, help, and the internal hook worker.Analysis does not need an API key. By default it reuses an LLM CLI already on PATH and logged in: claude, gemini, copilot, cursor-agent, or ollama. --in-chat instead asks the agent that just wrote the code to name the concepts, which is visible in the conversation and uses some of its context. Turns with no substantive code skip the LLM call. Failures in hooks are logged and swallowed so they cannot interrupt the coding session.The package is stdlib-only (Python 3.10 or newer) so hooks stay reliable in whatever environment an agent launches them. Install from PyPI with pip install learnlance. Internal hook and _worker subcommands are invoked by agent integrations, not by hand.

--in-chat

On setup / install: the agent analyzes its own work in the conversation instead of shelling out to a separate LLM CLI. Not available for git.

On setup: also print per-agent mechanisms and confidence. On list: include explanations.

Project directory instead of the current working directory (setup, install, uninstall, add use --path; show, list, stats, clear use --project).

Select a single agent on install or uninstall. Bare install detects all of them.

On show: write the HTML graph and do not open a browser.

On add: insert the concept from general knowledge even if no matching code is found.

On clear: skip the confirmation prompt.

On config: LLM CLI used for analysis (reads a prompt on stdin). Empty string restores auto-detection.

On config: optional model alias for the CLI backend (for example haiku).

On config: legacy explicit path to the claude executable. --llm-cmd supersedes this.

On config: pause or resume analysis without uninstalling hooks.

On config: run analysis in a detached process (on, default) or inline (off).

On config: maximum concepts extracted per agent turn (default 5).

~/.learnlance/config.json

Global settings. Override the directory with LEARNLANCE_HOME. Keys include enabled, llm_cmd (empty means auto-detect), cli_model, claude_bin, max_topics_per_turn, background, min_chars, and max_input_chars.

Per-project knowledge graph. graph.html next to it is the rendered view. ~/.learnlance/projects.json is the registry of known projects.

Per-session recap markdown.

Diagnostics used by doctor to tell configured hooks from hooks that have actually fired.

Alternative root instead of ~/.learnlance.

Set on LLM CLI processes spawned by learnlance so a Stop hook can ignore re-entrant calls.

setup

Detect supported coding agents and write their hooks for this project. Run this first after install. Same as a bare install.

Configure one agent by flag, or detect and configure all of them when no agent flag is given.

Remove selected agent hooks. With no flags, remove the Claude Code hook and disable auto-setup.

Print environment status, hook presence versus evidence of firing, LLM backend, and concept count.

Show or change settings (LLM command, model alias, enable/disable, background work, topic cap).

Render the project's knowledge graph to HTML and open it in a browser.

List learned concepts, most seen first. -v includes explanations.

Print concept, link, turn, and session counts, plus a per-category breakdown.

Search the codebase for a topic the agent missed, ask the LLM to explain it, and merge it into the graph.

Remove one matching concept (and orphaned related nodes), or wipe the entire graph when no concept is given.

Print the command menu.

Needs Python 3.10+ and, unless --in-chat is used, an LLM CLI already logged in. With no backend configured, edits are held rather than analyzed. --in-chat is not available for git, and it spends an extra agent turn that you see in the chat. A bare uninstall disables auto-setup until setup or install is run again. clear without -y refuses to wipe data when stdin is not a TTY. add still needs a working LLM CLI. Hook files follow each vendor's documented API but are unverified on a machine until doctor sees them fire. The HTML graph is local and offline; show opens it in the default browser unless --no-open is set.

learnlance is an MIT-licensed, stdlib-only Python CLI by aeroscissorz, first published in 2026. It is distributed on PyPI as learnlance, with the console script pointing at learnlance.cli:main.

claude(1), codex(1), copilot(1), gemini(1), ollama(1), git(1), pip(1)

Source code · Documentation