Confident AI

Native integrations post notifications to Slack and Discord through OAuth, to Microsoft Teams by webhook, to PagerDuty and to email, and let a person open Linear or GitHub Issues tickets from a trace; AI Connections call the customer's own AI endpoint to run evaluations, and connected MCP servers supply tool definitions so tool calls can be scored. These bring data in and send events out for the platform; none connects an agent to a real system to read, write or take actions there.

SourceConfident AI, confident-ai.com/docs project integrations, ai-connections and mcp-serversread 2026-09-21

Trace Workflows shows the pipeline that runs after a trace arrives (dataset ingestion, queue ingestion, evaluation rules and classifiers) as a graph, and the tracing SDKs record agent spans from the customer's own orchestration. No workflows that sequence, branch or retry steps, or combine deterministic nodes with agent steps, are documented.

SourceConfident AI, confident-ai.com/docs llm-tracing workflows and span-typesread 2026-09-21

Retrieval metrics such as contextual precision, contextual recall and faithfulness score the customer's own retrieval pipeline, retriever spans record the chunks that pipeline returned, and data source connectors feed documents into synthetic dataset generation. No index, retrieval layer or knowledge API that grounds an agent's behavior in company data is documented.

SourceConfident AI, confident-ai.com/docs metrics, span-types and data-source-connectorsread 2026-09-21

AI Governance (an Enterprise feature) groups controls into policies that are assessed daily and can block a release in CI/CD unless every control above Low importance passes, and each control resolves to a status that shows why it failed; annotation queues, fed automatically by queue ingestion tasks, custom annotation forms and end-user feedback put people in the review loop on traces and threads.

That gives a policy checkpoint on releases and a human review surface. No approval step or guardrail that holds or blocks an agent's own action in production is documented, and neither is a control to pause and resume a run.

SourceConfident AI, confident-ai.com/docs ai-governance, human-in-the-loop and threat-detectionread 2026-09-21

Self-serve SAML SSO works with Okta, Azure AD and Google Workspace; custom roles, policies and permissions apply at organization and project level, with all data separated by project; project and organization audit logs record every API and user action by user or API key; retention periods are configurable; data is encrypted at rest and in transit. The homepage states SOC 2 Type II compliance and that customer data is never used for training, and the data handling page states SOC II and HIPAA compliance with BAAs available.

SourceConfident AI, confident-ai.com/docs sso, rbac, audit-logs and data-handling, and confident-ai.com homepageread 2026-09-21

Tracing records traces, spans and threads, with typed spans for LLM calls, retrievers (the chunks returned), tools (which tool ran, with what arguments and what it returned) and agent runs that roll the nested steps up, plus logged prompts, token usage and cost. Project and organization audit logs record every API and user action separately from traces, traces export as CSV on demand or on a schedule to S3, trace forwarding sends every ingested trace to the customer's own OpenTelemetry collector, and retention periods are set per project or inherited from the organization.

SourceConfident AI, confident-ai.com/docs span-types, exports, trace-forwarding, audit-logs and data-retentionread 2026-09-21

Traces, threads and versioned datasets persist as observability and test records of what the customer's app did, and multi-turn metrics check whether that app retained knowledge across a conversation. No session, conversation, workflow or long-term memory that an agent reads and writes is documented.

SourceConfident AI, confident-ai.com/docs threads, dataset management and multi-turn metricsread 2026-09-21

The managed cloud stores and processes data in the United States by default or in the European Union when chosen at signup, the homepage lists regions in the US, EU, UK, Japan, Canada and Australia, and further residency is available on request. On Enterprise, self-hosting runs the whole platform inside the customer's own AWS, GCP or Azure account from a published Terraform module and Helm chart in an existing VPC or VNet, with guides for scaling, disaster recovery and air-gapped networks.

SourceConfident AI, confident-ai.com/docs self-hosting and data-residency, and confident-ai.com homepageread 2026-09-21

Confident AI publishes four official agent skills (confident-client, deepeval, confident-tracing and confident-otel) that teach a coding agent to build eval suites, instrument an app and administer the account, and ships them as ready-to-install plugins for Claude Code, Cursor and Codex; red teaming assessments come preset to OWASP Top 10 for LLMs and NIST AI RMF. These are packaged, editable starting points; no ready made agents, packaged employees or deployable workflows are documented.

SourceConfident AI, confident-ai.com/docs coding-agents skills and plugins, and confident-ai.com/llms.txtread 2026-09-21

Online evaluations run server-side on each trace and span as it is ingested, evaluation rules and ingestion tasks act on incoming data automatically, threshold alerts run on a recurring schedule and notify Slack, Discord, Microsoft Teams, PagerDuty or email, datasets can be evaluated on a schedule, and governance assessments run daily. All of it starts on an incoming event or a schedule, with no person asking.

SourceConfident AI, confident-ai.com/docs online-evals, workflows, alerts and ai-governance, and confident-ai.com/pricingread 2026-09-21

Each project chooses the provider and model behind its LLM-as-a-judge metrics, using the customer's own keys for OpenAI, Anthropic, Gemini, xAI, DeepSeek, Mistral or Perplexity, Amazon Bedrock or Vertex AI, or the Portkey and LiteLLM gateways, configured once at organization level and inherited by projects that opt in, or set per project; without customer keys, Confident-managed defaults apply. Model choice stays with the customer's admins, on their own providers and keys.

SourceConfident AI, confident-ai.com/docs evaluation-models and organization model-credentialsread 2026-09-21

A documented REST API covers datasets, test runs, traces, prompts, annotations, dashboards, governance policies and organization administration, with project and organization API keys that can be rotated and revoked; DeepEval is the open source Python SDK and confident-trace provides Python and TypeScript tracing SDKs, with OpenTelemetry ingestion from any language. Confident AI serves its own MCP server with OAuth dynamic registration for coding agents, publishes official agent skills and plugins, and fits CI through pytest and pull and merge request eval gates for GitHub and GitLab.

SourceConfident AI, confident-ai.com/docs api-reference, coding-agents/mcp, integrations and pr-eval-gate, and confident-ai.com/llms.txtread 2026-09-21

DeepEval and Confident AI test AI apps against versioned datasets of goldens with single-turn and multi-turn metrics, custom G-Eval and code-based metrics, and simulated multi-turn conversations; evals run as pytest unit tests in CI/CD with pass or fail thresholds, and the PR Eval Gate runs the app over a pinned dataset on every pull request and posts a passing or failing GitHub check. Online evaluations score production traces as they arrive, metric scores can be aligned against human labels, and production traces are curated back into datasets.

SourceConfident AI, confident-ai.com/docs llm-evaluation, unit-testing-cicd, pr-eval-gate and online-evalsread 2026-09-21

The Confident AI documentation covers evaluation, tracing, human review, red teaming, governance, settings and self-hosting, and no browser, desktop or computer control is documented.

SourceConfident AI, confident-ai.com/docs and llms-full.txtread 2026-09-21