coding-agents-kit
One command for every terminal coding agent: ak <kind> [@profile] — install, launch, switch accounts and read every account's quota, the same way for all of them.
$ ak usage
PROFILE PLAN WINDOW USED RESET
claude · Claude Code
@work max 5-hour ██▊░░░░░░░ 27% in 1h 15m
7-day all models ███▊░░░░░░ 38% in 1d 22h
@team team 5-hour ▋░░░░░░░░░ 6% in 3h 55m
codex
@work team 5-hour ████████▍░ 84% in 26m
7-day ██▌░░░░░░░ 25% in 5d 16h
antigravity · Antigravity CLI
@default AI Pro 5-hour Gemini ███▍░░░░░░ 34% in 1h 10m
7-day Claude and GPT ░░░░░░░░░░ 0% in 7d 0h
@work AI Pro 5-hour Gemini ▏░░░░░░░░░ 1% in 3h 55m
cursor · Cursor Agent
@default - n/d logged in: web session only
What it is
Every coding-agent CLI has its own flags for continuing a session, its own
way to keep a second account apart, its own headless mode and its own
"skip all permission prompts" switch. ak (also agentkit) puts one
grammar over all of them, so people, scripts and other agents drive any CLI
— with any of your accounts — without learning its quirks.
Supported coding agents
| Agent | Command | ak kind | Several accounts (@profile) | Quota in ak usage |
|---|
| Claude Code | claude | claude | yes — its own config dir | 5-hour and weekly windows, per model |
| Codex | codex | codex | yes — its own CODEX_HOME | 5-hour and weekly windows, resets held |
| Cursor Agent | agent | cursor | yes — its own home and credential store | connected only (web session) |
| OpenCode | opencode | opencode | yes — its own XDG dirs | through the provider key |
| Pi | pi | pi | yes — its own agent dir | through the provider key |
| Cline | cline | cline | yes — its own CLINE_DIR | through the provider key |
| Grok | grok | grok | yes — its own GROK_HOME | connected only (per-session usage) |
| Antigravity CLI | agy | antigravity | yes — its own home (one Google login each on Linux and in containers) | Gemini and Claude/GPT, 5-hour and weekly |
ak install <kind> installs each one from its vendor, pinned to an exact
version and verified (sha256 or npm integrity). Kinds are data in
providers.toml: adding an agent is a table, not code.
Provider routes
The same CLIs, routed to another model provider with one key per account:
| Provider | Key | Kinds | Quota |
|---|
| Z.AI GLM Coding Plan | ZAI_CODING_API_KEY[_SUFFIX] | claude-glm, codex-glm, opencode-glm, pi-glm | 5-hour and weekly credits |
| Azure OpenAI / Microsoft Foundry | AZURE_OPENAI_API_KEY[_SUFFIX] | codex-azure, opencode-azure, pi-azure, cline-azure | in the Azure portal |
| xAI | XAI_API_KEY[_SUFFIX] | codex-xai, opencode-xai, pi-xai, cline-xai, grok | needs a management key |
| Gemini API | GEMINI_API_KEY[_SUFFIX] | antigravity | in Google AI Studio |
What you get
| |
|---|
| One grammar | -c, -r [id], -l mean the same in every CLI; ak maps them to each one's own form (the grammar). |
| Several accounts | ak claude @work, ak codex @client-x: each profile has its own login, sessions, settings and keys, side by side. |
| Quota of every account | ak usage (or ak status) lists every account of every CLI in one table, ak pick names the one with the most room, ak capacity judges one. |
| Headless runs | ak run runs any kind non-interactively and returns one JSON result — the transport other tools and agents use. |
| Agents launching agents | ak pick --count N, leases and the agentkit-orchestrate skill spread lanes over accounts and vendors. |
| Autonomy by default | each CLI's own autonomy flag on every launch, with an opt-out that always wins: --ask or AGENTKIT_PERMISSIONS=ask. Meant for disposable or sandboxed environments. |
| Keys stay in one file | ~/.config/agentkit/env (mode 600); no key lands in a config file, a log or an output. |
| Set up by an agent | the agentkit-setup skill and the read-only ak setup probe | verify | docker install and repair the kit on any host or container. |
| Interface 1 for tools | ak run, ak env, ak doctor --json, with versioned JSON schemas. |
Requirements: bash (macOS, Linux, WSL) or PowerShell (Windows), and
python3 ≥ 3.9. Nothing else.
Install
git clone --branch v1.1.0 https://github.com/DailybotHQ/coding-agents-kit && ./coding-agents-kit/install.sh
Verify the checkout against the release's checksums (every shipped file is
listed):
cd coding-agents-kit
curl -fsSLO https://github.com/DailybotHQ/coding-agents-kit/releases/download/v1.1.0/SHA256SUMS
shasum -a 256 -c SHA256SUMS --ignore-missing # Linux: sha256sum -c SHA256SUMS --ignore-missing
Windows: .\coding-agents-kit\install.ps1. Scripts and containers:
./coding-agents-kit/install.sh --no-rc. Details, upgrades and uninstall:
docs/install.md.
Let an agent do it: npx --yes skills add DailybotHQ/[email protected] --skill agentkit-setup, then ask it
to set the kit up on this machine or to write a container recipe (setup).
Quickstart
ak doctor # what is installed and configured
ak install codex # a missing CLI, from its vendor, pinned
ak # list the kinds and which CLIs are installed
ak claude # Claude Code with its autonomy flag (the default)
ak codex -c # continue the newest Codex session (codex resume --last)
ak claude @work -c # a second Claude account, continue its newest session
ak claude-glm # Claude Code routed to Z.AI GLM (key from ~/.config/agentkit/env)
ak claude --ask # opt out for this launch: the CLI asks before acting
ak claude -- @src/app.ts # `--` sends everything after it to the CLI untouched
ak usage # quota and usage of every account, one row each
ak run codex --cwd ~/src/api --timeout 900 --output-format json -- "fix the failing test"
Several accounts
ak profiles add claude @work # a second Claude account
ak claude @work # log in AS the profile (the first launch asks)
ak usage # every account's windows, read-only
ak pick claude # the Claude account with the most room
ak capacity claude @work # one account: ok | low | critical | unknown
ak run claude @work --cwd ~/src/api --output-format json -- "fix the failing test"
A named profile reads its own keys (ZAI_CODING_API_KEY_WORK for
claude-glm @work) and never falls back to the default key. See
Accounts and Orchestration.
Choosing accounts, and agents launching agents
$ ak pick claude --count 2
claude @work2 headroom 100% binding 5h resets in 3h 48m
claude @work3 headroom 100% binding 5h resets in 3h 58m
ak pick claude --model fable --json # pick-v1: chosen + every account with its reason
ak profiles --json # profiles-v1: every account by name, no values, no paths
ak run claude @work2 --cwd ../wt-api --trust-cwd --output-format json -- 'Implement the API lane; commit.'
--trust-cwd marks the lane's folder as trusted in that profile first, so a
new account does not stop at its folder-trust screen (opt-in only). A Claude
profile can also log in with its own token, CLAUDE_CODE_OAUTH_TOKEN_<SUFFIX>
(from claude setup-token), for containers that reuse the host's accounts.
The loop — inventory, pick, one worktree per writer, failover, verify, merge
— is Orchestration.
Long runs stay ahead of their limits: an agent asks for its own room at each
task boundary and hands off before it runs dry, and leads lease the accounts
they hand out so two never share one Codex login:
ak capacity --self --json # capacity-v1, from a cache of usage numbers (ak usage stays live)
ak lease codex @work1 --owner local:w1:p2 # advisory; Codex takes one holder
ak pick codex --owner local:w1:p1 --count 2 # spreads lanes: every active lease counts, yours too
Skills for agents
| Skill | For | Install |
|---|
agentkit | launching kinds, accounts, ak run, ak usage, ak pick (ak --skill) | npx --yes skills add DailybotHQ/[email protected] --skill agentkit |
agentkit-setup | installing, configuring and repairing the kit (ak --skill setup) | npx --yes skills add DailybotHQ/[email protected] --skill agentkit-setup |
agentkit-orchestrate | several agents over several accounts and vendors (ak --skill orchestrate) | npx --yes skills add DailybotHQ/[email protected] --skill agentkit-orchestrate |
The grammar
The grammar is the same for every kind:
ak <kind> [@profile] [--ask | --auto] [-c | --continue | -r [id] | --resume [id] | -l] [--] [cli args…]
| Flag | claude | codex | cursor | opencode | pi | cline | grok | antigravity |
|---|
-c | --continue | resume --last | --continue | --continue | --continue | newest session here → --id <id> | --continue | --continue |
-r <id> | --resume <id> | resume <id> | --resume=<id> | --session <id> | --session <id> | --id <id> | --resume <id> | --conversation <id> |
-r | --resume (picker) | resume --all | --resume | refused: run opencode session list | --resume | history | --resume | refused: no picker (-r <conversation_id> or -c) |
-l | (passed through) | resume --last | ls | (passed through) | (passed through) | (passed through) | (passed through) | (passed through) |
Only the head of the arguments belongs to ak; the first token it does not
recognise, and everything after it, reaches the CLI unchanged. ak
replaces itself with the CLI (exec), so Herdr and your terminal see the
agent itself.
How it fits with DeepWorkPlan
ak works on its own. In a repository that runs DeepWorkPlan,
the pack's agentkit addon uses ak run as the headless delegation transport
for a plan's parallel_safe tasks, and the lead assigns one account per lane
with ak pick --count; interactive lanes run in Herdr panes through
herdr-peers, started as
ak <kind> @account --trust-cwd. devcontainer-kit
installs the kit, pinned and verified, in its images. The kit page is
deepworkplan.com/kit/agentkit.
Part of the DeepWorkPlan ecosystem
DeepWorkPlan turns any repository into an AI-first,
agent-pilotable codebase. These projects ship together; every addon is optional.
| Repository | Role | Purpose |
|---|
| deepworkplan-skill | Pack | The methodology, specification and kit as an installable agent skill |
| coding-agents-kit (this repository) | Addon: agentkit | One command for every coding agent: accounts, quota, ak pick, headless runs |
| devcontainer-kit | Addon: devcontainer | A reproducible dev container per repository, wired to Herdr and the host's accounts |
| herdr-peers | Addon: Herdr | Ask an agent in another Herdr pane, on any machine, and get one authorized reply |
| deepworkplan-vim | Addon: DeepWorkPlan Vim | The terminal editor for plans: plan browser, lanes view, Markdown viewer |
| ai-diff-reviewer | Required local review | Reviews a change set locally and in CI; every Final Review runs it |
| agent-skill | Addon: Dailybot | The Dailybot agent skill: progress reports, check-ins, chat and more |
| deepworkplan.com (source) | Site and hub | The methodology, specification, kit catalog and /init.md; coordinates the ecosystem |
Security
Autonomy by default with an always-effective opt-out (--ask,
AGENTKIT_PERMISSIONS=ask), keys only in ~/.config/agentkit/env (mode 600),
no network except ak install fetching a pinned vendor installer you asked
for and ak usage / ak pick querying the usage endpoint of an account you run (one
request per account, read-only). Report vulnerabilities privately — see SECURITY.md;
the threat model is docs/SECURITY.md.
Documentation
- Kinds and providers — the data model, provider keys, adding a kind
- Accounts — the account model: per-CLI table, lifecycle, keys, macOS and Windows, troubleshooting
- Profiles — accounts per CLI, keys per profile, Herdr hooks
- Permissions — autonomy by default, the
--ask opt-out, the classic and providers aliases
ak run — the headless contract, exit codes, the JSON envelope- Set up on any machine or container —
ak setup, the agentkit-setup skill, container recipes
ak usage, ak pick, ak capacity and ak lease — quota per account, reading the table, choosing an account, one account's room, the usage cache, leases (usage-v1, pick-v1, capacity-v1, leases-v1)- Orchestration — several agents over several accounts: worktrees, failover, Herdr, Deep Work Plans
Contributing
Contributions are welcome: CONTRIBUTING.md (setup, the
test gate, commit and PR flow) and AGENTS.md for coding agents.
Please follow the Code of Conduct.
License
MIT — see LICENSE. Credits in CREDITS.md.
Part of the DeepWorkPlan ecosystem — works on its own.
