Features¶
Everything Evokai does today. The README has the short list.
| ๐ง Any model | Bedrock, OpenAI, Azure, Mistral, Ollama and more through pydantic-ai, behind an allow-list, with an ordered fallback chain when a model has no capacity, and prompt caching per agent on Bedrock Anthropic models. |
| ๐ง Your tools over MCP | The app's MCP server, called as the signed-in user (forward_user) or with a service secret. Tools are filtered per user; a server that is down only loses its own tools. |
| ๐งช Evals | Test cases per agent (contains, patterns, JSON schemas, tools called, model-judged rubrics), run on demand and on every new version against the previous one, with their own budget. |
| ๐ Ready connectors | SQL databases, REST APIs from their OpenAPI document, Google Drive, SharePoint and OneDrive, IMAP mailboxes: enabled by configuration, read-only unless you opt in, under the same grants. |
| โ Confirm before change | Tools that change data never run directly: the turn stops with a confirm event and resumes on the user's decision. Questions with options work the same way. |
| ๐ Skills | Folders with a SKILL.md and, beside it, references, data and scripts: only names and descriptions go in the prompt, the agent loads a skill when it needs it (or it is preloaded), reads its files one at a time and runs its scripts in the code sandbox when code execution is on. |
| ๐๏ธ Voice | Dictation, conversation and barge-in. Browser engines (on-device recognition, in-browser Piper or device voices) or server voice on Pipecat over WebRTC, with Silero VAD and Smart Turn. |
| ๐ Rich answers | Charts, schemas, navigation and questions as typed events, rendered by your own interface. |
| ๐ถ Budgets | Hard caps per application, defaults per user (month and day), enforced before every model call. People, in the app and in the console, only ever see percentages of the caps, never amounts. |
| ๐๏ธ Your choice of storage | Conversations stored in Evokai (any SQL database) or kept by the app and sent with each request. A time to live per application deletes them for good. |
| ๐ก๏ธ Agent security layer | Prompt-injection and jailbreak heuristics, untrusted tool output and e-mails marked as data, personal data (Italian formats included) masked or tokenised, secrets and unsafe links stopped in answers, a tool policy, abuse limits; each control off, monitored or enforced, per application and per agent, with security events. Mapped to the OWASP LLM Top 10. |
| ๐ช๐บ Privacy | A person's data exported or erased on request, sign-out everywhere, retention per kind of data, and a data residency per application that keeps models in their geography. |
| ๐ Compliance packs | AI Act (classification, Annex IV-style documentation, changes), GDPR (record of processing, DPIA template, rights, DPA annex) and an ISO 27001 / SOC 2 mapping, generated from the live configuration in English and Italian, every statement with its source, snapshots and diffs. |
| ๐งฎ Code execution | Agents run Python for analysis and charts in Amazon Bedrock AgentCore Code Interpreter: one sandbox per call, no internet, EU region, output through the guard, off until granted. |
| ๐ Console access | Single sign-on (OpenID Connect, PKCE, MFA from the provider) with viewer, operator and owner roles enforced on every route, and an audit log of every change. |
| ๐ Never stored at providers | No provider-side threads, assistants or stored responses, ever. History is always sent from Evokai or the app. Enforced by a test. |
| ๐ Open standards | AG-UI for CopilotKit frontends, an OpenAI-compatible API for any OpenAI client or LangChain, and A2A agent cards for other agents, under the same keys, grants, guard and caps. |
| ๐ค Back-office agents | Agents that run server to server with a JSON output schema, validated, for automation such as e-mail triage. |
| ๐ช Admin console | Applications, agents with versions, per-application capability grants, MCP servers, skills, keys, limits, usage, a playground and health, at {base}/admin/. |
| ๐ญ Observable | Structured logs with correlation ids and per-call provider metrics; OpenTelemetry traces, metrics and logs when you turn them on. |