Most of a coding agent's context goes on finding things out: searching, opening files, following calls, checking what was tried before. All of it stays in the context and crowds out the work the agent was asked to do.
ask-kit moves that work out of the agent's context. It gives the agent one tool, ask. The
agent asks a question in plain words, and a cheap investigator works it out in its own context.
The agent gets back a short answer with the evidence it's based on, and its own context stays
free for the change it is making.
you: add a CSV exporter like the other ones
agent → ask: how are the exporters built and registered, and which tests cover them?
ask → Exporters implement Exporter in src/export/base.ts:12 and are registered in
src/export/index.ts:30. JsonExporter (src/export/json.ts) is the closest model;
tests are in test/export/*.test.ts.
(This is an example of what an answer looks like. Context, time and cost per ask has real numbers.)
A real ask from the shell, about this repository's code, with a line per program the investigator runs and then a follow-up:
ask is an MCP server, so any agent that takes MCP servers
can use it: Claude Code, Codex, GitHub Copilot CLI, Gemini CLI, opencode.
Read the guide: how to use it, use cases, models, tuning. Also read it on temet.ai.
For each step of a question, the investigator writes a small JavaScript program and runs it in a REPL. The program does whatever the step needs: loops, filters, joins, counts, comparisons, parsing. It is written fresh for that question. A domain gives the programs access to the data. What a program prints is all that returns to the investigator's model. Code mode, recursive language models and Claude Code subagents each do part of this; How ask-kit relates to them goes through each.
Each thing ask can answer about is a domain. ask-kit comes with two, as a starting point:
- code: the repository the agent is working in, cited as
path:line. - sessions: your past agent sessions, from the transcripts Claude Code, Codex, Copilot CLI, Gemini CLI and opencode already save. It reads them where they are, with no proxy or database.
Add your own for anything else your agent keeps digging through.
Paste one line into the agent you already use, on Linux, macOS or Windows:
Read https://github.com/temet-ai/ask-kit/blob/main/AGENTS.md, then set up ask-kit for me.
The agent installs ask-kit and asks you a few questions. The main one is which model does the digging:
- an API key (cheapest and fastest),
- a subscription you already pay for (Claude Code, Codex, Copilot),
- or a local model.
It shows what it will change before it changes anything. Restart your agents afterwards, and
they call ask by themselves when they need to find something out.
In Claude Code you can install it as a plugin instead. Its npm packages come with the install, and it needs Node.js 22.13 or newer:
/plugin marketplace add temet-ai/ask-kit
/plugin install ask-kit@temet-ai
The plugin brings the ask tool, the skills and the tuner agent. Then ask Claude to set up
ask-kit, and the ask-kit-setup skill asks you the same questions. The plugin covers Claude Code
only; ask-kit init and ask-kit setup leave Claude Code alone while it is installed.
In Gemini CLI, gemini extensions install https://github.com/temet-ai/ask-kit adds the ask
tool, run through npx @temet-ai/ask-kit. Run npx @temet-ai/ask-kit setup once to choose the
model.
Without an agent: npm install -g @temet-ai/ask-kit, then ask-kit init and ask-kit setup.
It needs Node.js 22.13 or newer, and git for the code domain's history and for shared kits. If npm answers EACCES, add
--prefix ~/.local to the install command.
Set it up has the details.
How much context one ask keeps out of the calling agent. These are medians from mike-ask, the proof of concept ask-kit grew from:
ask-kit itself, on 18 Nov 2026, on one machine with about 1,500 sessions on disk:
| Question | Investigator | Time | Cost |
|---|---|---|---|
| What did a Codex session find and change | glm-5.3-flash (OpenRouter) | 16 s | $0.0022 |
| Which sessions worked on X in the last week | glm-5.3-flash, domain picked by the router | 53 s | $0.0083 |
Follow-up on that (continue: true) |
glm-5.3-flash | 17 s | $0.0035 |
| Which Copilot session did a piece of work, under which ticket | Claude Code, haiku | 23 s | $0.0025 |
| same | Codex (ChatGPT plan) | 36 s | in the plan |
| same | Copilot CLI | 134 s | Copilot premium requests |
| How does routing work in this repo | deepseek-v4.1-flash (code) | 110 s | $0.0033 |
| Folder | What |
|---|---|
bin/, src/ |
the ask-kit command, the MCP server, the investigator loop, the transcript readers |
domains/ |
the two built-in domains, code and sessions: instructions, scripts and the typed description of what they can reach |
bundle/ |
what ask-kit init installs into each agent: three skills, the ask-kit-tuner agent and the when-to-ask rule. They are templates, and init fills in this machine's paths. |
docs/user-guide/ |
the guide, also at temet.ai/docs/ask-kit |
Early. Tested on Linux with real transcripts. Fixtures for all five transcript formats run in CI
on Linux, macOS and Windows. The Gemini CLI reader and the opencode and Gemini install formats
follow their docs and haven't yet been tried on a real install. Releases are on npm as
@temet-ai/ask-kit.
ask-kit started as mike-ask, a proof of concept that answers from what mike records of every Claude Code session. ask-kit is the general version. It needs no mike and no server.
It sends your question, and the excerpts the investigator reads, to the model you chose. There's
no telemetry. Everything it keeps is under ~/.config/ask-kit/ and ~/.local/state/ask-kit/,
including the clipped prompts ask-kit gaps caches from your transcripts and, for the one in 20
action answers that a model checks in the background, both answers.
What it stores and sends has the full list.
Bug reports and small fixes are welcome; see CONTRIBUTING.md. Report security problems privately, as SECURITY.md says.
MIT, see LICENSE. Third-party licences are in THIRD-PARTY-NOTICES.md.

