The StudioOverviewExtensionsHow It Works
ResourcesBlogFAQAbout
Get in touch
Specialist AI agents working your project as a team.
A GDPR expert, a backend architect, a tester, a design consultant — specialised AI agents working side by side on your project. Each with its own provider, identity, and perspective. Each scoped to its own provider, identity, and domain.
Every specialist gets their own workspace.
Tabbed multi-pane layout. Your security reviewer in one tab, your architect in another, your tester in a third — all seeing the same project, each through their own lens.
Your workspace at a glance — projects, agents, and recent activity in one place.
Task delegation is not the same as collaboration.
Claude Code can spawn sub-agents — copies of itself dispatching tasks in parallel. That's task dispatch, not team collaboration. A sub-agent won't flag a GDPR risk the coder missed, because it has the same knowledge, the same priorities, the same blind spots.
The Studio lets you run a security reviewer, a backend architect, a test engineer, and a compliance expert on the same project — each with its own model, its own instructions, its own domain knowledge. They see the same code from different angles.
Through the Orchestrator, they discuss, disagree, propose, and vote. A coder writes the migration. The architect reviews the schema. The GDPR expert flags the personal data column. The tester asks where the rollback plan is.
Data sovereignty built into every layer.
Every capability is designed around one principle: you own your data, your process, and your agent strategy. No vendor can take that away.
Define a security reviewer, a backend architect, a compliance expert, a test engineer — each with its own LLM provider, model, instructions, and domain knowledge. They work the same project from different perspectives, catching what a single agent would miss.
- Built-in profiles: backend, frontend, security, test, devops, architecture
- Custom roles with instructions, tone, constraints, and skills
- Each agent scoped to specific tools and file paths
- 6+ LLM providers — pick the right model for each role
Every agent authenticates with its own token. Every tool call is tagged with agent identity before it reaches the Orchestrator. Governance and audit trails work at the individual agent level, not the user level.
- Agent-specific auth headers on every MCP request
- Role templates with remit and boundary enforcement
- Local MCP router multiplexes identity across agents
- Ties directly into Orchestrator governance
Three-layer PII detection — regex (500+ patterns), entropy analysis, and ML classification — masks sensitive data before it reaches any LLM. A local vault restores originals for display. Your data never leaves unprotected.
- Emails, credit cards, tax IDs, API keys, 30+ country formats
- Entropy-based password and secret detection
- ML layer (Piiranha ONNX) for names and usernames
- Per-category toggles — you choose what gets masked
Conversations are encrypted client-side with AES-256-GCM before upload. Keys are derived from your password via Argon2id. The server stores ciphertext it cannot decrypt. Optional escrow for organisational data recovery.
- Argon2id key derivation (65536 memory, 3 iterations)
- Per-session Data Encryption Keys (DEK)
- Password cached in OS keychain for auto-unlock
- Bidirectional sync with conflict resolution
Five plugin types — CLI agents, MCP servers, daemons, exporters, utilities — with 15+ extension points. Built-in plugins handle code intelligence, documents, terminals, and application observation. Third parties use the same typed SDK.
- Full UI pages: config, status, workspace, tabs
- MCP tools callable by any agent in the workspace
- Terminal tools with streaming progress
- Marketplace for discovery and installation
Agents in the Studio aren't standalone. They participate in server-side conversations, vote on proposals, access shared knowledge, follow gated plans, and track tickets — all coordinated through the Orchestrator.
- Real-time state sync over WebSocket
- MCP tool calls routed through local gateway
- @-mentions inject knowledge, plans, tickets into chat
- Agents check remit before acting — governance enforced
Your data never leaves unprotected.
Before any message reaches an LLM, Data Shield scans it through three layers. Regex catches structured patterns — emails, credit cards, API keys. Entropy analysis finds passwords and high-entropy secrets. The ML layer (Piiranha) identifies names and usernames in natural text.
Detected PII is replaced with deterministic tokens. A local vault maps tokens back to originals for display — but the original values never reach the model.
When conversations sync to the server, they're encrypted client-side first. AES-256-GCM with Argon2id-derived keys. The server stores ciphertext it cannot decrypt. Your password never leaves your machine. Optional RSA-2048 escrow for organisational data recovery.
Five built-in extensions, open to third parties
Each extension runs in isolation with its own MCP tools, UI pages, and lifecycle. Third-party developers use the same typed SDK as built-in extensions.
A team that discusses, disagrees, and builds consensus.
A team of specialists that discuss, disagree, and build consensus. Connected to the Orchestrator for shared knowledge, gated plans, and governance. Your data never leaves unprotected.