Several AI specialists, one focused workspace.
Agent Studio gives each assistant its own system prompt, model, tools and private memory. Link your PDFs, approve web searches when fresh information matters, and switch between specialists without losing the thread.
Prompts, memory and documents stay with the agent.
A general chatbot forgets what it is for. Here an agent is a durable configuration: the prompt you wrote, the model and style you picked, the PDFs you linked, and the conversation so far. Add a second agent and it starts clean, with its own thread and its own sources.
- Per-agent prompts, history and linked documents
- Durable LangGraph threads with a SQLite checkpointer
- Local multilingual embeddings over your own PDFs
- Human approval before any web search leaves the workspace
Prompts are first-class
Start from a template, write your own, or let Mistral draft a structured prompt from a one-line brief.
Two models, three styles
Mistral for multilingual work and vision, GPT-OSS for heavier reasoning, each with creative, precise or fast settings.
Retrieval with sources
PDFs are chunked, embedded locally with e5-small and filtered by document id, so answers can point back to a page.
Guardrails by default
Input limits, per-account rate limits, PII redaction and sanitized output run before and after every model call.
Durable threads, not a chat log
Each agent runs on a LangGraph thread backed by a SQLite checkpointer. State is saved per step, so a restart resumes the same conversation. It is durable execution in miniature; Temporal adds distributed workers, retries and schedules at a scale SQLite cannot hold.
A short tour
Four screens, no sign-up: chat, prompt editing, document library and live retrieval.
Look around first.
The demo is open, read-only and rate limited. The full workspace is invite-only and uses Google sign-in, so there is no shared password to hand out.