Agent ForgeSelf-hostable, multi-tenant AI agent platform

RBAC · Observability · Quota management · Admin dashboard Pluggable agent framework adapters · OpenAI-compatible LLM streaming

Get Started

View on GitHub

Agent Forge — Gateway, Harness, Runtime, Telemetry, LLM API

Agents — pluggable runtime

One AgentRuntime interface, many frameworks. DirectLLM (DeepSeek) ships today; Google ADK and LangGraph adapters plug in with zero gateway changes.

Skills

Reusable skill packages across three layers — user, project, and per-workspace. RBAC-guarded, audited CRUD wired straight into the harness.

MCP integrations

Workspace-scoped MCP server registry with CRUD, live tool discovery, and health checks — extend agents with external tools and context.

Global usage control

Tenant-wide quota management with token and cost tracking. QuotaGuardrail enforces limits per workspace and feeds the usage dashboards.

Multi-tenant + RBAC

Workspace-scoped RBAC driven by permissions.yaml — from viewer to tenant_admin, with scoped admin access and a single source of truth.

Observability

Traces, metrics, and quota signals exported via OpenTelemetry — live dashboards for requests, sessions, cost, errors, and SSE events.

How it works

A thin API Gateway routes authenticated, authorized requests to a pluggable Agent Runtime — governance and observability wrap every execution.

Gateway

FastAPI handles JWT auth, RBAC permission checks, audit logging, and quota enforcement.

Harness

Guardrail pipeline and sandbox wrap tool execution with retry and circuit breaking.

Runtime

An AgentRuntime adapter calls the LLM, streaming events back over SSE.

Telemetry

Spans, metrics, logs, and quota usage are exported via OpenTelemetry.

quickstart.sh

# Start the API gateway
uvicorn src.main:create_app --reload --port 8000

# Stream a chat (SSE)
curl -N http://localhost:8000/api/v1/chat \
  -H "Authorization: Bearer $TOKEN" \
  -d '{"agent":"assistant","message":"hello"}'

3architectural layers

4RBAC roles

7build phases

404tests