Agent Dot
Give your agents a real job.
Most agent frameworks produce impressive demos and fragile products. Agent Dot is the infrastructure layer for agents that hold a real job: it gives them playbooks they can load on demand, memory that survives sessions and can be audited like code, sandboxes where they can safely execute work, and connectors to the systems your business already runs on. The output is artifacts, reports, code, presentations, not just text.
Why agents fail in production
Six failure modes kill most agent deployments: prompts that break across model updates, memory that vanishes between sessions, sandboxes that leak, MCP servers managed by hand, skill routing that guesses wrong, and hard vendor lock-in. Agent Dot was designed against this exact list, each subsystem exists because one of these failures took down something we were running.
Artifacts, not answers
An agent with a real job produces things you can ship: a formatted report, a working pull request, a finished presentation. Agent Dot's sandbox and skill system are built around producing and persisting artifacts, files live in workspaces, survive the session, and are yours.
Run it your way
Deploy as a self-hosted Docker Compose stack or on your cloud. You hold the encryption keys, the data, and the skill library. Teams comfortable with Git and Docker are productive in an afternoon.
What's in the box.
Skills as playbooks
Markdown playbooks with triggers, workflows, and references that agents load on demand. Update one file, every agent improves, no retraining.
Git-versioned memory
User, session, and project memory with full provenance. Inspectable, revertable, never silently overwritten, deletions cascade cleanly.
Sandboxed execution
Per-session Docker containers with bind-mounted workspaces and an allowlist package proxy. Agents run real code without touching the host.
MCP connectors
PostgreSQL, Slack, GitHub and more, configured once with encrypted secrets, pooled across sessions, activated with an @mention.
Three-stage skill routing
Keyword scan, embedding search, then LLM tie-break. Requests reach the right playbook even when user intent drifts mid-conversation.
Write once, swap models
Claude, OpenAI, Gemini, Ollama, and Azure Foundry on an identical codebase. Your agent logic outlives any single provider.
Works with
ClaudeOpenAIGeminiOllamaAzure FoundryDockerMCPGitHub