01 · Live gallery

Agent Fleet

A self-hostable multi-agent operations platform — chat with a fleet of tool-using agents, hand the orchestrator a goal and watch it execute as a live task DAG, with evals, cost governance, and guardrails.

Agent Fleet is a self-hostable multi-agent operations platform. You chat with a roster of tool-using agents, or hand the orchestrator a goal and watch it decompose into a live task DAG with human-approval gates.

The public repository is https://github.com/hharsha98/agentfleet. The product spine is FastAPI plus Next.js, Postgres with pgvector, and a hand-built agent runtime with an env-switchable LangGraph path. The fleet can consume external MCP servers and also expose itself as an MCP server for IDE clients.

Ops is part of the product: per-message metering, cost budgets, prompt-injection screening, PII masking, versioned agent publish/rollback, and an Eval Center with a CI regression gate. It is meant to be run locally or self-hosted, not presented as a hosted SaaS. The public gallery is a click-through of the product surfaces.

In the repo

  • Streaming multi-agent chat with per-agent tools and prompts
  • Orchestrator: goal → async DAG with human-approval gates
  • Document RAG: local embeddings (fastembed) + pgvector citations
  • MCP in both directions — consume servers and expose the fleet
  • Eval Center, cost budgets, guardrails, versioned rollback

Stack

PythonFastAPINext.jsPostgreSQLpgvectorMCPDocker