AI providers and models

Every AI feature in kcap can run on the managed Kurrent AI provider with nothing configured, or on a provider you bring. This page is where each feature’s switch lives

Kurrent AI, and what your plan gets

Titles, outcome summaries, embeddings and server evaluations run on Kurrent AI on every plan, including Free. The Insights agent, work items and projects need Team or Enterprise, and bringing your own key does not unlock them. On the local edition there is no Kurrent AI; bring keys or run without

Providers: AI & models → AI Providers

One credential per vendor, shared by evaluations, Insights, titles, summaries and the work-items agent:

Vendor	Needs	Also
Anthropic	API key
OpenAI	API key	optional Base URL for a gateway or compatible service, including the /v1 path
Cohere	API key, or a base URL for an on-premises deployment
OpenRouter	API key	models as provider/model

Max Parallel Judges per vendor (1–16, default 4) bounds concurrent evaluation calls. A stored secret stays when its field is blank; Clear on save removes it. Model pricing shows when prices last synced from models.dev; cost estimates depend on it

Evaluations: Evals → Evals

• Evals agent → Provider: Auto (Kurrent AI preferred) uses Kurrent AI when available, else the first configured provider in the order Anthropic, OpenAI, OpenRouter, Cohere. An explicit choice is a pin and does not fall back
• Eval judge: the model per vendor. Defaults: claude-sonnet-4-6, gpt-5-mini, command-a-03-2025, google/gemini-2.5-flash
• Run limits: run timeout (default 30 minutes), per-question judge timeout (5 minutes), time reserved for the retrospective (2 minutes), and the trace size substituted into a prompt (400,000 characters). A question timing out does not stop the others
• Judge-fact deduplication: embeds findings and merges near-duplicates into clusters; needs an embedding provider; enabling it backfills earlier findings

Turn automatic evaluations on

Per-repository settings: Auto-evaluate on close queues an evaluation when a session ends, and queues the ended sessions not yet evaluated; Promotion floor sets the minimum weight for a finding to be promoted. On plans with projects, Per-project settings governs every repository in a confirmed project, shown as via project on their rows. Scopes without an explicit choice default to on when Kurrent AI is the effective provider. These switches save immediately

Evals → Eval Queue shows waiting, in-flight, failed and cancelled server runs, one at a time with manual runs first. Breakers show a vendor-wide failure such as bad credentials or exhausted quota; fix the cause, then Reset. Dead letter holds runs that exhausted retries, with Re-enqueue. Evals → Eval Catalog edits the questions, the shared template (which must contain {QUESTION_TEXT}) and the retrospective prompt, with version history

Embeddings: AI & models → Embeddings

One provider for memories, fact clustering and work-item matching: Voyage (default voyage-4-lite), OpenAI (text-embedding-3-small, optional dimensions and base URL), Kurrent AI, or on the local edition Ollama (mxbai-embed-large). Changing provider, model, dimensions or host changes the embedding identity: stored vectors are discarded and retained facts re-embedded in the background, which can cost provider usage. A credential-only change does not

Titles and outcomes: AI & models → Summaries

Session titles: Auto prefers a single tenant key for Anthropic, OpenAI or OpenRouter, then Kurrent AI, then Cohere; or pick a provider. Per-vendor models; OpenAI needs an explicit one. Auto-backfill titles ended sessions that still have a raw first-message title, off by default, and sends session content, including private sessions, to the provider. Imported sessions are titled on arrival regardless

What’s Done: the outcome summary. Same as session title inherits the title provider, or choose one

Insights: AI & models → Insights

Team and Enterprise. From kcap 1.4 the agent is on by default where a provider is available; earlier versions start off. Choose a provider and model, or Default for the automatic choice (Anthropic when available, else Kurrent AI, else the first configured). Row limit and query timeout are adjustable

Flows: Flows → Flows and Flow Catalog

Flows → Flows sets the limits for dynamic flows, the one-off definitions an agent supplies when it starts a run: an Enabled switch (off blocks new dynamic starts and new rounds, without ending a round in progress), the maximum participants (default 3), simultaneous dynamic runs per person (2), rounds (10), spend in USD (5), the round timeout cap and default (15 and 10 minutes), the idle lifetime cap (2 hours), and an optional model allowlist. Catalog definitions are exempt from these clamps. A dynamic participant must name a concrete, priced model

The Claude reviewer certification panel can require a daemon’s launcher policy version and restrict the Claude CLI versions allowed to run unattended reviews, or block Claude reviewers outright; leave both version fields blank to skip the gate

Flows → Flow Catalog is the YAML editor for named definitions, with validation, version history, Enabled, Delete and Restore; the format and an example are in Flows

Work items: Work Items → Work Items

The correlation agent: Enabled, provider (Auto as for evaluations) and an optional model. Scope: a confirmed project’s switch governs its repositories; other repositories have their own; defaults are on when Kurrent AI is the effective provider. Turning a scope on backfills its last 30 days of unassigned sessions. Work item operations lists durable merges, resolutions and undos that need attention, with Accept and Retry