EDDI Conversational AI Orchestration
Run configurable multi-agent workflows that connect conversations, models, tools, and business systems in a self-hosted service.
- Source repo
- labsai/EDDI
- Stars
- ★ 383
- Last updated
- today
- License
- Apache-2.0
- Primary language
- Java
- FA score
- 68/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI APIClaude API (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- An enterprise platform team routing requests to specialized agents can define intent and behavior rules, then deploy agents through Manager.
- Not a fit if
- Teams seeking a vendor-hosted SaaS with no service deployment
- Users unwilling to configure Docker and a database
- Source review
- 68/100 · Some gaps
What does this agent do, and when should you use it?
EDDI (Enhanced Dialog Driven Interface) is self-hosted conversational AI orchestration middleware built with Java 25 and Quarkus, with agent behavior defined in versioned JSON configurations. It routes user requests to agents, which can call LLMs, external APIs, and MCP tools, while MongoDB or PostgreSQL stores conversations and persistent memory. The service includes a Manager dashboard, a Workforce workspace for group conversations, and a standalone Chat UI; it also exposes REST, MCP, A2A, and the OpenAI Chat Completions API. Its installer starts EDDI and a selected database with Docker Compose, and production deployment options include Kubernetes and OpenShift. It fits teams seeking a governable, self-hosted multi-agent service, with model endpoints or API credentials configured separately for actual inference.
A user submits a request through the built-in Chat UI, a REST endpoint, or an integration, and EDDI routes it according to configured behavior rules. An agent runs its configured workflow and can call an LLM, HTTP APIs, built-in tools, or external MCP servers; group conversations let multiple agents discuss, vote, or work through phased tasks. EDDI stores conversation and user-memory data and can retrieve documents through RAG; it supports file uploads, website crawling, and asynchronous REST document ingestion. Operators configure, test, and deploy agents in Manager and review logs, approvals, usage, and audit records. Deployed agents can also be exposed to compatible clients through the OpenAI Chat Completions API, MCP Server, and A2A interfaces.
- An enterprise platform team routing requests to specialized agents can define intent and behavior rules, then deploy agents through Manager.
- A business team seeking peer review or structured decisions can configure group discussions with Peer Review, voting, or Task Force phases.
- A knowledge team answering questions from internal documents or websites can build RAG knowledge bases and schedule ingestion.
- An integration engineer connecting agents to existing APIs and tools can configure HTTP tools, connections, an MCP Client, or an agent from an OpenAPI specification.
- A regulated team requiring review gates and traceability can use human approval workflows, the audit ledger, and data export and erasure features.
How do you install or deploy this agent?
On Linux, macOS, or WSL2, install Docker. The installer uses Docker Compose to deploy EDDI and the selected database, and generates a unique vault encryption key. After startup, open http://localhost:7070 and choose Manager or Workforce; the standalone chat UI is at /chat.
curl -fsSL https://raw.githubusercontent.com/labsai/EDDI/main/install.sh | bashWindows PowerShell:
Invoke-WebRequest -UseBasicParsing -Uri "https://raw.githubusercontent.com/labsai/EDDI/main/install.ps1" -OutFile "install.ps1"
Unblock-File .\install.ps1
.\install.ps1The installer supports options such as PostgreSQL and Keycloak:
bash install.sh --db=postgres --with-authBuilding from source requires JDK 25; the repository includes a Maven Wrapper. The development setup also requires MongoDB 6.0+, while Docker is used for integration tests and container builds.
How do you use this agent?
After startup, open Manager (/manage) to create an agent with the wizard or configuration editor. Set its model provider, workflow, and tools, deploy it, then try it in /chat. Model credentials depend on the selected provider; the vault requires a master key, which the installer generates. The provided Ollama Compose configuration can run a local model.
The installer creates a CLI wrapper for updates:
eddi updateDeployed agents can be called by compatible clients through the OpenAI-compatible API, which is disabled by default. MCP clients can connect to /mcp; on an authenticated instance, they must sign in through OAuth.
What are this agent's strengths and limitations?
- Agent behavior lives in versioned JSON, so configurations can change without redeploying the application.
- Supports 19 LLM providers, model cascading, MCP, A2A, and an OpenAI-compatible API.
- Includes MongoDB and PostgreSQL persistence options, RAG, cross-session memory, a management UI, and audit features.
- Provides human approval gates, a secrets vault, SSRF protections, and data export and erasure capabilities.
- It requires deploying a Java/Quarkus service and a database; the quick install uses Docker, and production needs authentication and key configuration.
- LLM endpoints and credentials must be configured; cloud model calls may incur separate usage charges.
- Source development requires JDK 25; full integration tests require Docker, and Kubernetes deployment requires additional secret and database setup.
- Documented upgrades include cross-version database and Keycloak adjustments, so existing deployments may need manual migration work.
How does this agent compare with similar options?
The README compares EDDI with LangGraph, CrewAI, and AutoGen, describing EDDI as focused on JSON configuration, Java 25 virtual threads, and production governance. EDDI suits teams wanting a separately deployed service and UI for managing agents; teams that prefer orchestration embedded directly in application code can evaluate those library frameworks.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| EDDI Conversational AI Orchestration This agent | 68 · Some gaps | Self-hosted serviceFree + model costs | ★ 383 | today | Java | OpenAI API |
| Agent Swarm | 47 · Major gaps | Agent plugin / skillFree + model costs | ★ 873 | today | TypeScript | Codex · Claude Code |
| CrewAI Studio | 51 · Major gaps | Web appFree + model costs | ★ 1.4k | 2mo ago | Python | OpenAI API · Claude API |
| Flow-Like | 55 · Major gaps | Desktop appFreemium | ★ 961 | today | Rust | — |
How does FollowAgents rate this agent?
Why each dimension lost points
The evidence describes a secrets vault, URL validation rules, a prohibition on dynamic code execution, MCP/OAuth authentication guidance, and human approval for Platform Operator writes. The auto-approval workflow also limits permissions, checks CI and review conditions, and excludes workflow changes. These concrete controls support moderate to strong scores, but runtime implementation details are absent, so equivalent safeguards across all tools and external calls cannot be confirmed. Rollback is sparsely described. Publisher identity is unverified, and the supplied repository material does not clearly establish the maintainer, so attribution is only partial.
The README describes version following, uncommitted failed-task output, retries, and dead-letter tracking. Supplied tests show regression coverage for mock-route shadowing and asynchronous toast cleanup, while security and dependency update processes are documented. Scores are limited because this material cannot establish end-to-end Agent reliability; the test count and coverage are README claims, and comprehensive failure semantics or dependency availability guarantees are not shown.
The materials identify concrete scenarios including support routing, group discussions, team tasks, RAG, Slack, OpenAPI, MCP, and local or cloud models. They also document database, authentication, monitoring, GPU, and local LLM deployment choices, supporting strong scenario and environment-fit scores. Configuration constraints, approval gates, and some streaming fallback cases describe boundaries, but tool permissions, model capabilities, and all degradation boundaries are not fully specified.
The README has a substantial table of contents, quick start, deployment options, upgrade guidance, linked documentation, troubleshooting notes, and examples. The license file explicitly provides Apache 2.0. Deductions reflect limited naming and version evidence (for example, Java 25 is stated, but the supplied material lacks a corresponding release history), limitations being distributed across documentation, and maintenance responsibility not being inferable from unverified publisher identity. The security policy supplies supported versions, a reporting address, and response timelines, but does not establish a specific accountable maintainer.
If implemented as described, configurable orchestration, model and protocol integrations, memory, and RAG could reduce glue code and ease deployment; several interface and API forms are described. Scores remain modest because the supplied evidence provides no independently verifiable outcomes, cost data, or assessment of configuration complexity against benefits. Production-grade and compliance statements are primarily repository claims.
The README links many capabilities to specific guides, the security policy and workflows expose inspectable controls and update steps, and the test files explain concrete test purposes. This supports partial traceability and corroboration across source types. Deductions reflect that major capability and compliance statements remain largely README assertions without the relevant implementation files or external validation; the supplied evidence does not independently substantiate badge and test-count claims.
- This review uses only the supplied static materials and has low confidence; README feature, coverage, and compliance claims are not independently verified here.
- Before enabling external tools, LLMs, OAuth, or remote access in production, check deployment settings, secret permissions, and human-approval boundaries. The README says authentication should be enabled when binding services beyond localhost.
- The security policy lists 6.0.x and 5.6.x as supported; the supplied materials do not identify the release version corresponding to this revision.
FAQ
Does EDDI cost money to use?
Do I have to use a hosted LLM?
Where do I create my first agent?
http://localhost:7070, go to Manager, and use the agent wizard or configuration editor to create and deploy an agent.Can EDDI act as an MCP tool server?
/mcp is an OAuth-protected resource, so MCP clients must sign in.