clikernel is a tiny stdin/stdout worker around execnb.shell.CaptureShell. It keeps one IPython-compatible Python process alive and returns concise text for each request.
The idea of clikernel is to be a persistent python process that LLMs can use as their primary tool.
A persistent Python process is a good default tool for an LLM agent. The agent can import a module, inspect it, keep helper functions around, cache results, patch objects, and then continue from that state on the next request.
This is especially useful with pyskills. The agent can discover skills, read their docs, import the ones it needs, and keep using them in the same session. That makes Python a universal workbench for repo inspection, API checks, file edits, data transforms, and experiments.
clikernel gives an agent the part of a notebook kernel it usually needs: send code, wait for the result, read concise text, and keep the Python state.
On startup, clikernel prints loading status followed by a random session delimiter:
please wait, loading...
loading complete. first delimiter:
--aB3x9
That delimiter stays the same until the worker exits.
Send one line to execute it immediately:
1+1
Each complete request is acknowledged with . before execution starts. Send exit() or quit() to receive an acknowledgement, a final delimiter, and stop the worker.
For multiline code, send -- on its own line, then the code, then the session delimiter exactly:
--
def f(x):
return x + 1
f(2)
--aB3x9
After execution, clikernel prints the acknowledgement, the rendered output, and the session delimiter:
.
3
--aB3x9
Outputs are rendered with fastcore.nbio.render_text. A single non-empty output is printed directly. Multiple outputs use raw XML-ish tags, for example <stdout>, <display_data mime="text/markdown">, and <execute_result>.
clikernel registers two line magics wrapping execnb's nbopen/nbrun, for running cells from a notebook by cell id prefix:
%nbopen foo.ipynb
%nbrun ab12
%nbrun ab12 --above
%nbrun --all --exported
%nbrun ab12 --fname other.ipynb
%nbopen sets the default notebook for later %nbrun calls (as does passing --fname). %nbrun runs the cell whose id starts with the given prefix; --above/--below also run the cells before/after it, --all runs every code cell, and --exported filters to cells with an nbdev #| export/#| exports directive. The notebook is re-read from disk on each call, and each executed cell's rendered output is printed under a --- {cell id} --- header.
clikernel sets quiet defaults for IPYTHONDIR, MPLCONFIGDIR, and MPLBACKEND=Agg before creating the shell. Existing IPYTHONDIR and MPLCONFIGDIR values are left alone. Loading messages and any startup warnings are printed before the first delimiter. Set CLIKERNEL_STATE_DIR to choose the default parent directory.
clikernel can check each cell before it runs, to warn about or forbid certain code. On startup it loads inspectors from $XDG_CONFIG_HOME/clikernel/inspectors.py (usually ~/.config/clikernel/inspectors.py). If that file is absent, nothing changes.
Each cell is transformed first (so IPython magics and ! shell escapes parse), and its AST is passed to every inspector before the cell executes. An inspector returns a string to prepend a note to the cell's output, raises to block the cell (it does not run, and the exception is reported), or returns None to do nothing. Define a function inspect(tree) and/or a list inspectors of such functions in the file. A broken inspectors.py is reported on stderr and skipped, so it cannot stop the kernel starting.
See examples/inspectors.py for one that blocks subprocess, os.system/os.popen, and ! escapes, steering the agent to the permission-checked Bash tool instead.
clikernel is built for a client that reads stdout as tokens. Local echo is disabled when stdin is a TTY. The client already knows the code it sent, so echoing it back only makes the LLM read slow, expensive tokens that add no information.
Each complete request prints . on its own line before execution starts. That gives the client a cheap early byte to read, which is useful when the request will run for a while. Each response ends with the same delimiter on its own line. The client can read until that line appears instead of parsing prompts or waiting and guessing. The delimiter is random per process, so it is unlikely to appear in generated code, copied logs, examples, or earlier transcript text. Keeping it fixed for the session means a client does not get stuck just because it missed a rotated delimiter.
Startup messages appear before the first delimiter. After that first delimiter, the stream follows the request-response protocol. Outputs are rendered as concise text, using unescaped XML if required when there's multiple outputs.
IPython history is disabled. IPYTHONDIR and MPLCONFIGDIR get quiet defaults when the environment has not already set them, and MPLBACKEND defaults to Agg.
pip install -e .[dev]Version lives in clikernel/__init__.py as __version__.
- Ensure your GitHub issues are labeled (
bug,enhancement,breaking). - Run:
ship-gh
ship-pypi
ship-bump # dev release always later than prod release