CLI reference

amber login

Sign in through your browser and store an API key for one Workspace. Shows a code and opens /cli/authorize; once you approve, the CLI stores the key.

--api-url <url>

Where Amber is (default: https://amber.trycoredepth.com).

--no-browser

Print the link instead of opening it.

--json

Print JSON instead of a table, for scripts.

amber logout

Forget the stored API key on this machine. The key stays valid until you revoke it in Workspace settings.

--json

Print JSON instead of a table, for scripts.

amber whoami

Show who you are signed in as and which Workspace the key is for.

--json

Print JSON instead of a table, for scripts.

amber version

Print the installed amber-cli version.

--json

Print JSON instead of a table, for scripts.

amber benchmarks list

List the public Benchmarks, or your Workspace’s private ones.

--private

Your Workspace’s private Benchmarks, not the public ones.

--limit <n>

Show at most this many (default: all).

--json

Print JSON instead of a table, for scripts.

amber benchmarks show <slug>

Show a Benchmark: what it measures, its versions, and what a Job on it typically costs.

--json

Print JSON instead of a table, for scripts.

amber models list

List Models with their Providers and prices (USD per million tokens).

--provider <slug>

Only Models served by this Provider, e.g. anthropic.

--limit <n>

Show at most this many (default: all).

--json

Print JSON instead of a table, for scripts.

amber models configs list

List Model Configurations: a Model, its pinned Provider and settings. The slug is what -m takes.

--model <slug>

Only this Model’s, by slug.

--provider <slug>

Only those pinned to this Provider.

--limit <n>

Show at most this many (default: all).

--json

Print JSON instead of a table, for scripts.

amber quote

Price a Job without starting it. Shows the breakdown (model, sandbox, judge, fee) and how confident the estimate is. Nothing is reserved until you amber run. A Quote leaves out combinations that don’t fit.

-b, --benchmark <slug[@label]>

Benchmark, e.g. terminal-bench@2.0 or gsm8k. Without a label, the newest version that isn’t deprecated. Repeat for several.

-m, --model <slug>

Model Configuration slug, e.g. opus-high. Repeat for several.

-a, --agent <slug>

Agent, e.g. terminus-2. Repeat for several. Without it, the default Agent is used, and the Direct Agent too for a Benchmark with static Tasks; with it, exactly the Agents you name.

-n, --trials <1-10>

Trials per Task. Defaults to 1.

--task-id <id>

Only this Task of the chosen Benchmarks. Repeat for several; each Benchmark needs at least one.

--json

Print JSON instead of a table, for scripts.

amber quote -b terminal-bench@2.0 -b gsm8k -m opus-high -m gpt-high -a terminus-2 -n 1

amber run

Quote a Job, ask to confirm, start it and follow its progress. It spends money, so it never does so silently: anything but a person at a terminal (a script, a pipe, --json) must pass --yes, or the command exits 2 without spending.

-b, --benchmark <slug[@label]>

Benchmark, e.g. terminal-bench@2.0 or gsm8k. Without a label, the newest version that isn’t deprecated. Repeat for several.

-m, --model <slug>

Model Configuration slug, e.g. opus-high. Repeat for several.

-a, --agent <slug>

Agent, e.g. terminus-2. Repeat for several. Without it, the default Agent is used, and the Direct Agent too for a Benchmark with static Tasks; with it, exactly the Agents you name.

-n, --trials <1-10>

Trials per Task. Defaults to 1.

--task-id <id>

Only this Task of the chosen Benchmarks. Repeat for several; each Benchmark needs at least one.

-y, --yes

Don’t ask for confirmation.

--name <name>

A name for the Job.

--detach

Start the Job and return without following it.

--json

Print JSON instead of a table, for scripts.

amber run -b terminal-bench@2.0 -m opus-high -m gpt-high -n 3

amber jobs list

List recent Jobs in the current Workspace.

--state <state>

Only Jobs in this state: running, completed, cancelled…

--limit <n>

Show at most this many. Defaults to 20.

--json

Print JSON instead of a table, for scripts.

amber jobs show <job-id>

Show a Job’s state, spend and progress.

--json

Print JSON instead of a table, for scripts.

amber jobs cancel <job-id>

Stop a Job. Running Trials stop now and are charged for what they used; the rest of the Quote goes back to your Wallet.

--json

Print JSON instead of a table, for scripts.

amber jobs wait <job-id>

Follow a Job until it ends. Exits 0 only if it completed (6 otherwise).

--json

Print JSON instead of a table, for scripts.

amber results <job-id>

Show a Job’s Rewards and costs per cell (Benchmark × Agent × Model Configuration).

--trials

Also list every Trial.

--json

Print JSON instead of a table, for scripts.

amber compare <job-id>...

Compare Model Configurations that are comparable (the same Benchmark Version and Agent Version), with 95% intervals and “too close to call” flags. Others are listed separately, with the reason.

--benchmark-version <id>

Benchmark Version id, if the Jobs have several.

--agent-version <id>

Agent Version id, if the Jobs have several.

--json

Print JSON instead of a table, for scripts.

amber export <job-id>

Export a Job’s results. The file is written atomically; without -o it goes to standard output.

-f, --format <json|csv|html>

json (full detail), csv (one row per Trial) or html (a report). Defaults to json.

-o, --output <file>

Write to this file (default: standard output).

--json

Print {path, bytes, format} instead of a table. Needs -o.

amber export 7f3a --format html -o report.html

amber wallet

Show your Wallet: what it holds, what is set aside for open Jobs, and what you can spend.

--ledger <n>

Also show the last n Wallet entries.

--json

Print JSON instead of a table, for scripts.

amber upload <path>

Upload Harbor Tasks as a private Benchmark: a folder of Task folders, or a .tar.gz or .zip of them. Validation is quoted first, then run; the Tasks that passed are published. Like run, it asks before spending.

--name <name>

Name for the Benchmark (default: the folder’s name, or with --benchmark that Benchmark’s).

--benchmark <slug>

Upload the next version of this private Benchmark.

--allow-host <host>

A host the Tasks’ Environments may reach besides the Amber model proxy (repeatable).

-y, --yes

Don’t ask for confirmation.

--exclude-failing

Publish the Tasks that pass if some fail.

--no-publish

Stop after validation.

--label <label>

Label for the new version (default: the next number).

--json

Print JSON instead of a table, for scripts.

amber upload ./my-tasks/ --name my-bench