Kortix (Suna)
also known as Suna, Kortix Suna, Company AI Command Center
Type: full-code · Vendor: Kortix · Language: TypeScript-dominant monorepo (web, api, cli, desktop, sandbox) · Status: active · Status in practice: emerging
Open-source, self-hostable company command center (~20k GitHub stars): a workforce of specialist agents works in disposable per-session sandboxes on git branches, lands changes only through human-approved change requests, and runs on demand, human-assisted, or on schedules and triggers.
Description. Kortix (repo still named suna after the original Suna agent) treats the company itself as a git repository: agents, skills, integrations, automations, and memory are versioned files, and the platform 'opens pull requests against itself'. Every session runs in its own disposable Linux sandbox on its own branch; only committed work survives, and work reaches main solely through a reviewed change request. Skills capture reusable know-how shared into every session, connectors broker 3,000+ apps plus MCP/OpenAPI/GraphQL server-side through one scoped token, secrets are encrypted and injected at runtime without model exposure, and cron or webhook triggers spawn sessions automatically. Positioning is enterprise-grade: microVM isolation, per-resource permissions for people and agents, audit trail, human approval gates, and on-prem or air-gapped deployment.
Agent loop shape. Project (git repo + kortix.yaml) -> session spawns an isolated cloud sandbox on a session branch -> agent (OpenCode) works, commits, pushes -> change request -> human reviews and merges to main; thousands of sessions can run in parallel on the same config.
Primary use cases
- org-scale agent workforce producing reviewed deliverables (decks, reports, code, replies)
- scheduled and trigger-driven automations end to end
- self-hosted / air-gapped enterprise agent deployment
Key concepts
- Company as a git repository (docs) — Agents, skills, learned facts, and infrastructure definitions live as versioned, diffable, owned files.
- Change request gate (docs) — Work reaches main only through a human-approved change request; the company self-improves one reviewed change at a time.
- Skills (docs) — Reusable know-how encoding how the company does a job, written once and shared into every session.
- Triggers (docs) — Cron and signed webhooks that spawn sessions automatically.
Patterns this full-code implements —
- ★★Sandbox Isolation
Per-session disposable sandboxes on dedicated branches; only committed work survives.
- ★★Human-in-the-Loop
All agent work lands through reviewed change requests; approval gates guard sensitive actions.
- ★★Tool Use
Connector layer brokers thousands of apps and protocol surfaces server-side.
- ★Skill Library
Skills are first-class platform objects shared into every session.
- ★★Scheduled Agent
Cron and webhook triggers spawn sessions; automated work runs end to end on schedules.
- ★Agent Credential Vault
Secrets are scoped, encrypted, injected into sandboxes at runtime, and never exposed to the model.
- ★★Cross-Session Memory
A shared company memory compounds what agents learn across sessions; today file-based.
Neighbourhood
Click any neighbour to follow the lineage. Scroll to zoom, drag to pan.
Alternatives & relatives
References
Provenance
- Last analyzed:
- Last updated:
- Verification status: partial