See how Swarmix works

A 52px sidebar with icons for Home, Projects, Agents, Plans, Knowledge, Artifacts, Marketplace, Analytics, and Settings.

Each project binds to a local folder. The context panel shows project name, path, active agents, and session history.

Each agent slot combines: vendor + model, profile (persona + constraints), role (permissions + tool access), blocks, kit (MCP tools), and permission mode.

Choose from Plan (agent asks before every action), Auto-edit (files yes, commands no), or YOLO (full autonomy). Org policy can restrict which modes are available.

Every message, tool invocation, and result is recorded. Tool groups are collapsible. You see exactly what the agent did and can approve or reject actions.

A hybrid regex + ML model (Piiranha, ~200MB on-device) scans outgoing messages for PII. Detected fields are masked before reaching the LLM. Runs locally, nothing leaves the device.

An AI-driven controller that decomposes the user prompt into a multi-step plan, routes each step to the best-fit agent, and manages inter-agent handoffs.

Before an agent passes work to another, a gate validates the handoff: is the output complete? Does the receiving agent have the right tools? The user can require manual approval at each gate.

Ship with archetypes like Coder, Architect, Security Auditor, Technical Writer. Each includes a persona, skill triggers, and engineering constraints.

Roles can explicitly block tools (e.g., deny bash for a read-only reviewer). The client enforces these before the LLM even sees the tool. Denied tools cannot be overridden by the agent.

Agents share findings on a conversation board. Proposals require consensus voting from team members before proceeding. The governance layer enforces remit boundaries.

Structured plans with phases, items, and gate validations. Tickets track bugs and action items. Both link to conversation threads for full context.

An MCP stdio tool that provides read/write/convert operations for office formats. Agents can extract tables from Excel, read PDF content, and generate Word documents.

Drag and drop files into the chat. The agent processes them locally. Large files are chunked. Results stay on your machine unless you explicitly share them.

The Knowledge Intelligence plugin can automatically extract key decisions and findings from conversation transcripts into draft knowledge articles for review.

Articles move through Draft → Review → Approved → Deprecated. Each stage has ownership, ratings, and full edit history. Promoted articles become the team reference.

OpenTelemetry auto-instrumentation captures tool invocation latency, active session count, and error rates. Export to any OTLP-compatible backend for dashboards and alerting.

Each conversation session shows a chronological view of messages, tool calls, approvals, and cost. Filter by agent, project, or date range.

The PolicySettingsEnforcer on each client merges org rules with local config. Blocked features are disabled in the UI. Policy violations are logged and can trigger alerts.

Enterprise IdPs (Okta, Azure AD, OneLogin) sync users and groups via SCIM 2.0. Deprovisioning can soft-deactivate or hard-remove membership, sessions, and project access.

Electron desktop app. Runs on the user's machine. Handles chat UI, local file access, and agent subprocess management. No inbound network ports.

Organization-level server. Manages org config, user membership, policy distribution, and SSO/SCIM integration. The single source of truth for org policy.

Docker Compose for single-VM. Helm chart for Kubernetes. PostgreSQL for state. All config via environment variables. Health endpoints for load balancer probes.

Clients connect to Hive over WSS (port 443). Hive connects to Orchestrator internally. No direct client-to-orchestrator traffic. All traffic encrypted in transit.

Org scope comes from the SCIM bearer token, never URL params. External IDs are per-org. Provisioning in org A never creates membership in org B. Seven hard isolation rules enforced.

Tested against Okta, Azure AD, and OneLogin SCIM clients. Supports bulk operations (up to 100 ops), path-less PATCH (Azure AD), and standard filter syntax.

Google and Microsoft OAuth supported out of the box. JWT tokens carry org context. Session revocation is immediate.

Org admins generate and revoke SCIM tokens, view token health alerts, and configure deprovisioning strategy per org.

Org Admin sets policy in Hive. Hive sidecar distributes to clients. PolicySettingsEnforcer merges org rules with local config. PolicyFeatureGuard disables blocked UI elements. Violations are logged.

Block YOLO mode, restrict model selection (e.g., only Claude Sonnet), enforce Data Shield, control external MCP server connections, limit file system access patterns.

Regex catches structured patterns (SSN, credit cards, emails). The ML model catches unstructured PII (names in context, addresses, medical terms). Both run before the message leaves the client.

Org policy can make Data Shield mandatory. When enforced, users cannot disable it. The shield indicator is always visible in the chat header.

Active MCP sessions (gauge), tool invocation latency (histogram), tool invocations total (counter). SCIM-specific: provisions, deprovisions, auth failures, request duration, group ops.

Set OTEL_ENABLED=true and point OTEL_EXPORTER_OTLP_ENDPOINT to Grafana, Jaeger, Datadog, or any OTLP collector. Zero code changes.

Deactivate (default): soft-delete membership, user can be re-provisioned later. Full Remove: remove from all SCIM groups, delete project memberships, delete org membership, revoke all sessions.

All provisioning and deprovisioning events are audit-logged with timestamps, actor, and affected resources. OTEL metrics track provision/deprovision counts by org.

Projects bind to a local directory. The agent's file read/write/edit tools are scoped to this path. Git status and recent commits are available in context.

Resume interrupted sessions. The SessionResumeDialog shows previous sessions with their last message and context, so you can pick up where you left off.

Each agent slot has a kit: the set of MCP tools it can invoke. Add tools from the marketplace or configure custom MCP servers. Tools run as stdio subprocesses.

Plan: agent asks before every action. Auto-edit: file edits allowed, shell commands require approval. YOLO: full autonomy. Choose per agent.

Related tool calls collapse into groups (e.g., "5 file reads"). Expand to see details. Each call shows the tool name, arguments, result, and timing.

Files created or modified by the agent appear as artifacts linked to the conversation. Browse, diff, and manage them from the artifacts panel.

The conductor analyzes your prompt and generates a step-by-step plan. Each step is assigned to an agent based on its role and tools. You can edit the plan before execution starts.

When agent A completes its step, a handoff gate validates the output before passing to agent B. Gates check completeness, tool availability, and can require your approval.

Plans contain phases. Phases contain items. Items have status (pending, active, completed, blocked), acceptance criteria, and linked artifacts. Gates validate phase completion.

When work diverges from the plan, record a diversion with the reason. Memos park out-of-scope items for later. Full version history tracks every plan change.

Tickets have status transitions, priority, labels, and assignees. Dependencies track blockers. Subtasks break large tickets into manageable pieces.

Link tickets to external trackers (Jira, Linear, GitHub Issues). Pull latest state from the tracker. Two-way sync keeps both systems current.

The Knowledge Intelligence plugin scans conversation transcripts and surfaces candidate knowledge articles. Accept, edit, or discard extractions. Good ones become team reference.

Knowledge articles link to each other and to conversations, tickets, and plan items. The context view shows all related docs and links for any entry.

Each plugin lives in plugins/name/ with its own package.json, migrations, tools, routes, and optional frontend. Plugins import only from @swarmix/core, never from server or other plugins.

Core services communicate with plugins through a generic event broadcast. Plugins receive events (document created, proposal resolved, ticket transitioned) and react independently. Failures are isolated.

Organization → Projects → Agents. Each level has its own access controls. Org-level policies cascade to all projects and agents below.

Org admins manage user membership, configure SCIM, set deprovisioning strategy, define policy rules, and monitor compliance. All admin actions are audit-logged.

Agents call gov_check_remit before acting on topics. If the topic is outside their remit, they escalate via gov_escalate to involve the right agent.

The gov_responsibilities tool shows which agent owns which domain. Clear ownership prevents overlapping work and ensures accountability.

Layer 1 (Regex): catches structured patterns with known formats. Layer 2 (ML): Piiranha model catches context-dependent PII that regex misses. Both run before the outgoing message.

Org policy can make Data Shield mandatory. When enforced, the shield cannot be disabled by the user. The chat header always shows shield status. Violations are logged.

Org Policy (broadest): blocks features and models. Role: restricts tools, file patterns, and actions. Permission Mode (narrowest): Plan, Auto-edit, or YOLO. Each layer can only restrict, never widen.

Roles can explicitly deny specific tools. The client enforces denied tools before the LLM sees them. Even in YOLO mode, denied tools cannot be called. This is the hard boundary.

The transcript captures: who sent each message (user or agent), which tools were called with what arguments, what each tool returned, whether the user approved or rejected, and the final result.

OpenTelemetry captures tool invocation latency, error rates, active sessions, and SCIM operations. All metrics carry org_id labels for per-tenant dashboards.

Soft-deletes the org membership (deleted_at timestamp). User can be re-provisioned later. Group memberships and project access are preserved but inactive.

Hard-deletes: removes from all SCIM groups, deletes project memberships, deletes org membership, revokes all active sessions. Immediate and complete. Audit log retained.

Draft: author writes and edits freely. In Review: submitted for peer review. Approved: officially blessed, searchable reference. Deprecated: marked outdated, replaced by newer content.

Full edit history on every document. Ownership transfers are logged. Ratings and quality audits provide ongoing quality signals. The kb_audit tool runs quality checks.