Krun agent skill
- The official Krun agent skill teaches Claude Code, Codex and other coding agents to design Krun integrations: primitive selection, uncertainty handling, multimodal inputs and latency. Install it with
claude plugin marketplace add krun-ai/skillsandclaude plugin install krun@krun-ai, or withnpx skills add krun-ai/skills --skill krun. - No API or SDK changes.
Krun One V1: multimodal inputs and multi
- Krun One V1 (
krun-one-v1) is the new default model. Requests that pinkrun-one-v0orkrun-one-v0.3keep working: both ids are aliases ofkrun-one-v1. See Models. - Multimodal inputs.
contextcan be a string, as before, or a list of content parts:text,image,documentandaudio. Media parts reference files uploaded with the new/v1/assetsendpoints, kept for 24 hours. multi, a fourth question type that selects every option that applies, with an independent probability per option.- Pricing is unchanged. Media inputs are counted as input tokens, like text.
- Cold starts can take up to about 2 minutes. The API now waits up to 150 seconds for the model. Set client timeouts of at least 180 seconds, the default in SDK 0.3.0, and retry
504 UPSTREAM_TIMEOUT. See Timeouts and cold starts. - Python SDK 0.3.0 (
pip install -U krun-ai) adds content parts,multi,client.assetsand the 180-second default timeout. - TypeScript SDK:
@krun-ai/sdkis on npm (npm install @krun-ai/sdk). V1 multimodal support requires 0.3.0 or later.
Data controls and typed feedback in the Console
- Data controls. Organization owners can choose whether Krun may retain selected requests and feedback to improve its models. Off by default.
- Console usage shows decisions by primitive (choice, noul, score), API latency percentiles and errors by code.
- Playground feedback asks for the expected option (
choice), true / false (noul) or the expected level (score). - No API or SDK changes.
Krun One v0.3: choice, noul and score
- Decision primitives. Besides
choice, questions can now benoul(the probability that a yes/no proposition holds) andscore(an expected level and a distribution over ordered levels). Types can be mixed in one call. See Decision primitives. - Krun One v0.3 (
krun-one-v0.3) is the new default model.choiceanswers keep v0’s accuracy and format; requests that pinkrun-one-v0keep working. - Feedback accepts a typed
expectedvalue (choiceoption,noultrue/false,scorelevel).expected_decisionstill works. See Feedback. - Pricing is unchanged: one question is one decision, whatever its type, and usage still reports input tokens only.
- Python SDK 0.2.0 adds
NoulQuestion/ScoreQuestion, typedNoulAnswer/ScoreAnswerandresult.choice()/.noul()/.score(). The TypeScript SDK (coming soon) will ship with the same types.
Krun One v0
- Krun One v0 (
krun-one-v0), the first Krun decision model, for intent and tool routing. See Models and Benchmarks. POST /v1/decidewith the questions and answers contract: one context, 1 to 16 namedchoicequestions, one answer per question withchoice,probabilities,confidence,abstainandabstention_status. See Questions and answers.POST /v1/feedbackto report whether an answer was correct, per question. See Feedback.GET /v1/modelsto list available models.- Python SDK 0.1.0:
krun-aion PyPI, withKrunandAsyncKrunclients. See Python SDK. - TypeScript SDK: coming soon. See TypeScript SDK.