Agentkit addon
Every terminal coding agent has its own flags for continuing a session, its own way to keep a second account apart, its own headless mode and its own switch for skipping permission prompts. coding-agents-kit puts one command surface over all of them: ak <kind> [@profile].
This addon integrates the kit into DWP v7 (pack v7.9.0) as the headless delegation transport. It is optional: without it, every task runs in the current session, exactly as before. The kit itself is an MIT product that works without Deep Work Plan.
What the kit gives you
- One grammar for every CLI.
ak claude,ak codex,ak cursor,ak opencode,ak pi,ak cline,ak grokandak antigravity, plus provider routes (Z.AI GLM, Azure OpenAI, xAI and the Gemini API), with the same session flags:-ccontinues,-r <id>resumes. - Profiles.
ak claude @workruns a second account in its own home, kept apart from the first. - Headless runs.
ak run <kind> -- "<prompt>"runs one prompt non-interactively and returns a documented exit code, optionally as one JSON object. - A doctor.
ak doctor --jsonreports which CLIs are installed, the profiles, and the names of the keys that are set — never their values. - Verified installs.
ak install <cli>installs a missing CLI from its vendor’s official channel at a pinned version, checked against a pinned sha256 or the npm registry’s integrity. - Familiar names. Two alias presets, off until you turn them on:
classic(claudex,codexx,cursorx,opencodex,pix,clinex,grokx) andproviders(claude-glm,codex-azure,codex-xai,pi-glm, …), each oneak <kind>.
Accounts, quota and choosing an account
The kit treats every login or key you hold as an account: one CLI × one profile. Since v1.0.1, every profile — @default included — is an independent account in its own directory under the profiles root, preventing state leakage and credential collisions. ak profiles lists them with their login state and the names of their keys; ak profiles --json gives tools the same list (schema profiles-v1). Since v0.5.1 it reports where Claude Code’s token really is — file, keychain (macOS), none or unknown — by presence only, never reading the secret, and a Cursor login on Linux, in a container or on a Linux host, is recognised. Since v0.5.2, AGENTKIT_DEFAULT_PROFILE_<CLI> names the account a bare ak <cli> uses (a positional @profile still wins), which is how a dev container picks the host account it shares. Since v0.5.3, ak codex adds --no-daemon when a Codex home shared with another machine records a background daemon that is not on this one.
ak usage (also ak status) lists every account of every CLI the kit knows, one row per account and profile, under its CLI: live quota for Claude Code subscriptions (5-hour, weekly and per-model windows), Codex ChatGPT logins, Z.AI coding-plan keys and, since v0.7.0, Antigravity Google logins on Linux and in containers (the five-hour and weekly windows of agy’s own /usage panel). An account whose quota cannot be read still has its row: n/d with how it is connected and why — Cursor, Grok, xAI, Azure OpenAI, a Gemini API key, or an Antigravity login kept in the macOS Keychain — or no login with the command that logs it in, so the table also shows which accounts of each CLI are connected. It is read-only: it never refreshes a login, never prints a token and never writes under a profile. --json gives tools the same list (schema usage-v1). The usage endpoints are vendor-internal and unofficial.
ak pick <kind> reads the same rows and names the account with the most room. --count N returns up to N accounts, best first; --exclude @<profile> keeps a busy one out; --model counts one model’s weekly window; --need 5h:40 refuses to pick without that much room. --json gives the schema pick-v1.
ak usage # quota of every account and profile
ak pick claude --count 3 # the three Claude accounts with the most room
ak claude @work --trust-cwd # a second account, folder pre-trusted
ak profiles --json # profiles-v1, for tools
ak setup verify # ordered checks, each with its repair command
--trust-cwd (or AGENTKIT_TRUST_CWD=1) pre-trusts the folder the agent starts in, so Claude Code and Codex open without their folder-trust screen. ak setup probe | verify | docker is a read-only helper, driven by the agentkit-setup skill, that inspects the machine, checks the install and renders a pinned, verified container recipe.
Several agents, several accounts
The agentkit-orchestrate skill (ak --skill orchestrate) teaches an agent to spread work over several agents, accounts and vendors: one ak usage before the batch, ak pick --count N to choose the accounts, one delegate per account, the repository’s own gate on every result, and failover to the next account when one hits a limit. The kit observes, computes and runs; it never schedules. The lead — you, or an agent you asked — decides, using only your own accounts and within each vendor’s terms.
For long, interactive lanes, each lane runs in a Herdr pane on its own account. The methodology chapter Running many agents describes the whole model: a lead and its lanes, accounts chosen by headroom, herdr-peers coordination, and one branch and one release per repository; Swarm autonomy adds capacity checks, handoff and lead succession.
Capacity and leases
From v0.6.0 an agent can ask how much room its own account has, cheaply, at every task boundary — the signal that lets a swarm hand work on before an account runs out (Swarm autonomy).
ak capacity --self --json(schemacapacity-v1) gives the running account’sheadroom_pct, itsbinding_window,resets_atand averdict:ok,low(25 % or less),critical(10 % or less, or rate-limited) orunknown(no quota source: Cursor, Grok, or Antigravity on macOS). It usesak pick’s arithmetic, so the two never disagree, and reads a usage cache kept for 300 seconds (--max-age), so a check at every task boundary stays cheap.--selfresolves the accountaklaunched; outside anaklaunch it exits 3 and never guesses.ak usagestill reads live and refreshes the cache.- Thresholds and drills.
AGENTKIT_CAPACITY_LOWandAGENTKIT_CAPACITY_CRITICAL(or--lowand--critical) move the thresholds.AGENTKIT_CAPACITY_CAP=@<profile>=<pct>can only lower a reading — it is how a drill makes an account look poor — and can never make one look richer. ak lease <kind> @<profile> --owner <addr>holds an account for one lane;ak lease release …frees it once a successor has accepted the handoff, andak lease lslists the leases (schemaleases-v1). A Codex profile takes a single lease — one writer per auth file — and only the owner releases a lease. Leases expire and are advisory records of numbers, profile names and pane addresses, never a credential.ak pick <kind> --owner <lead>counts every active lease, the owner’s own included: it ranks candidates byheadroom / (1 + leases)and never offers a leased Codex account, not even to its holder.--reuse @<profile>hands one of the owner’s own leases on, for example to a lane’s successor. Without--ownerthe ranking is unchanged.
ak capacity --self --json # this agent's own room and verdict
ak lease claude @work --owner local:w2:p1 # hold an account for a lane
ak pick codex --owner local:w1:p1 --count 2 # skip accounts leased to others
Work moves only among the accounts you declared for the plan, within each vendor’s terms: no agent creates an account, renews a login or spends a Codex reset. The Deep Work Plan pack reads the verdict through its own helper and turns it into continue, prepare or hand off; without the kit, or with an older one, capacity is reported as unknown, never as ok.
Install
git clone --branch v1.1.0 https://github.com/DailybotHQ/coding-agents-kit && ./coding-agents-kit/install.sh
ak doctor
Requirements: bash on macOS or Linux, and python3 3.9 or newer; nothing else. Windows uses install.ps1. Pin v1.1.0: v0.2.0 and v0.2.1 are unsupported. Verify a release with its SHA256SUMS asset.
| Item | Value |
|---|---|
| Product | DailybotHQ/coding-agents-kit, tag v1.1.0, interface 1 |
| Registry key | agentkit in .dwp/config.json |
| Transport | headless: one ak run per delegate in a dedicated git worktree |
| Provides | subagents, cancel_children, model_routing |
| Requires | the plan contract’s agent_delegation grant |
Stable since v1.0.0. The kit’s public interface follows semantic versioning and changes incompatibly only in a new major version: the command surface (ak <kind> [@profile] for every kind in providers.toml, and every verb with its documented flags), the permission posture, interface 1 (ak doctor --json, ak env, the ak run grammar and envelope), the JSON schemas (doctor-v1, usage-v1, pick-v1, capacity-v1, leases-v1, profiles-v1, setup-probe-v1, setup-verify-v1), and the installed layout and env-file variable names. v1.0.0 changes no command, flag, key or schema from v0.7.1.
Autonomy by default, with an opt-out that always wins
Since v0.2.0, ak <kind> launches every agent in autonomy: it adds the CLI’s own autonomy flag, kept only in the kit’s providers.toml. Autonomy is meant for disposable or sandboxed environments, such as a development container.
The opt-out always wins: --ask on one command, or AGENTKIT_PERMISSIONS=ask in the environment or the kit’s env file, suppresses the flag even when the same command says --auto. An opted-out session passes the opt-out on to the agents it starts. On a host, set the opt-out.
The addon spells no autonomy flag and never passes --auto. It passes --ask when a plan records the opt-out. A plan that grants agent_delegation on a host accepts autonomous delegates confined to their own worktree, which is not a sandbox.
What it adds to a plan
On a v7 plan whose contract grants agent_delegation, execute may hand a parallel_safe task to another CLI: it creates a dedicated git worktree, runs ak run there with a timeout, and collects the result into the plan’s analysis_results/delegations/. The result is asserted evidence until the plan’s own gate runner observes it. Cancelling a delegate stops its whole process tree.
Agentkit or Herdr
| Situation | Use |
|---|---|
A bounded parallel_safe task with a declared output |
Agentkit (headless) |
| The task needs interaction, runs long, or lives on another machine | Herdr (a peer in a pane) |
The two compose: herdr-peers can launch a peer in a pane with the environment that ak env <kind> @profile prints. For many lanes at once, see Running many agents.
Notes
Optional and never required. API key values are never printed, logged or written to a config file; the documented exception is Cline, which receives its key on the command line. Result extraction for OpenCode, Pi, Cline and Grok is built from vendor documentation and has not yet been exercised against live accounts; unknown output falls back to raw text.