M
MCP Directory CURATED DIRECTORY
911 MCP servers indexed
中文
← Back to directory
A

Agent Swarm

Official
A multi-agent orchestration operating system for companies.
How to install ↓ Compare GitHub source repository ↗
Category
Dev Tools #199 of 438
Stars
★ 875 Very popular
Transport
Streamable HTTP
Runtime
Node.js · Docker
Credentials
API key / credential required
License
MIT
Last commit
54FMRS · D

The supplied material identifies Agent Swarm as an open-source, desplega-ai-maintained multi-agent orchestration system using streamable HTTP MCP transport. It supports local SQLite storage, Docker workers, shared memory, workflows, and multiple task integrations. The manifest does not list specific MCP tool names or client compatibility, so those fields remain empty.

Strongest · Documentation 14/20 Weakest · Reliability 8/20

Reliability
8/20
Security and permissions
9/20
Maintenance
12/20
Documentation
14/20
Setup experience
11/20
Why each score
Reliability 8/20
The manifest specifies an identifiable npm package, streamable HTTP transport, authentication headers, and a port, while the README describes the API/MCP service and several operating modes, so a normal initialization and tool-list flow is plausible. However, this is a static review with no source code, CI, or committed tests proving MCP handshake behavior, tool fidelity, error handling, or dependency control. The manifest default port 3000 also conflicts with the README's API port 3013, so substantial points are deducted.
Security and permissions 9/20
The API key is supplied through a required secret environment variable and an Authorization Bearer header, remote transport uses HTTPS, and the materials describe local-instance and agent-ID boundaries. No malware, credential theft, real secrets, or other red-line issue is evidenced. However, the system can orchestrate Docker workers, persistent memory, external integrations, scripts, and third-party routes, creating a broad side-effect surface. There is no source evidence for least privilege, tenant isolation, dangerous-operation confirmation, auditing, or data-flow controls, so the score remains moderate.
Maintenance 12/20
The repository is not archived and provides an MIT license, an identifiable maintainer organization, version 1.129.0, documentation, and contribution paths. The README also claims daily evolution, which is a positive maintenance signal. Still, there is no reviewable commit history, release cadence, dependency-update evidence, CI status, vulnerability-response process, or explicit security contact. The issue count alone does not demonstrate response quality, so points are deducted.
Documentation 14/20
The README covers architecture, use cases, integrations, CLI, deployment, Docker, environment-variable entry points, workflows, and many linked guides. It includes onboarding, npm/npx, and Docker examples, providing relatively broad layered coverage. It does not provide reviewable per-tool parameters, permission boundaries, limits, cost information, troubleshooting, or complete mainstream-client configuration evidence. The port examples also conflict with the manifest, and many capabilities remain unverified claims, so points are deducted.
Setup experience 11/20
The project offers bunx/npx onboarding, manual Docker Compose setup, API/worker/lead commands, and local endpoints, making a working instance plausible after runtime preparation. It still requires Docker, at least one harness credential, environment-file editing, and potentially additional integration configuration. No complete mainstream MCP-client connection example is supplied, remote deployment guidance is summarized rather than operationally complete, and the 3000/3013 port mismatch adds fragility, so the score is moderate.

Static review · not runListed 2026-08-14

Read the FMRS scoring method →

Fit and risk

What it can accessRuns commands or codeUses the network

Best for

  • Teams seeking self-hosted multi-agent orchestration
  • Engineering and operations teams that want a lead agent to coordinate specialized workers
  • Organizations needing shared memory, persistent identity, workflows, or human approval gates
  • Users able to manage Docker, API keys, and agent-harness credentials

Not for

  • Users who only need a single MCP tool or simple chat assistant
  • Environments that cannot run Docker or manage agent-harness credentials
  • Teams that do not want tasks to flow among multiple agents, shared memory, or external integrations
  • Users requiring documented compatibility with a ready-made desktop client

Required permissions

  • Access to the local or remote Agent Swarm MCP endpoint
  • An Agent Swarm API key supplied through an Authorization header with the Bearer prefix
  • The X-Agent-ID of a registered agent
  • Depending on enabled integrations and harnesses, credentials for Slack, code hosting, email, Linear, Jira, or other services

Risks and side effects

  • Worker agents execute tasks in isolated Docker environments but may access the code, files, and services configured for them
  • Shared memory retains learnings across sessions and can broaden the context visible to later tasks
  • Enabled integrations can cause messages, emails, ticket updates, code changes, or pull requests
  • Misconfigured API keys, Bearer headers, or agent IDs can cause connection failures or access to the wrong agent instance
  • Automated workflows and external actions should be reviewed with human approval gates, especially for irreversible operations

Setup

Before you start

Runtime：Node.js · Docker

AGENT_SWARM_API_KEY requiredsecret API key for your local Agent Swarm instance, used as the Bearer token; obtained from your deployed Agent Swarm instance and required for all requests.

Provide Docker and credentials for at least one supported agent harness. Run @desplega.ai/agent-swarm onboard with bunx or npx, or configure Docker Compose as described in the repository. After starting the MCP service, connect with the instance URL, a Bearer API key, and the X-Agent-ID of a registered agent.

Check that it works

Start the local API, then point your MCP client at http://localhost:3000/mcp with the Authorization: Bearer <AGENT_SWARM_API_KEY> and X-Agent-ID headers; if the connection succeeds and the Agent Swarm MCP tools (memory, KV, pages, etc.) appear in your client's tool list, the install works.

Troubleshooting

  1. Confirm that the Agent Swarm instance is running and check the port used in the MCP URL; the manifest defaults to 3000 while the quick-start example uses 3013
  2. Confirm that AGENT_SWARM_API_KEY is set and that the Authorization header contains Bearer followed by the key
  3. Confirm that X-Agent-ID belongs to a registered agent in the target instance
  4. Confirm that Docker is installed and running and that credentials for the selected agent harness are configured
  5. Use http://localhost:{port}/mcp for a local connection; for a remote connection, replace it with your own swarm host
  6. Check that the instance, port, API key, and agent ID match between the client configuration and the running server

Things to try

Once connected, you can ask your AI assistant things like:

  • Break down this feature request from Slack into subtasks, delegate them to workers, and ship the result as a pull request.
  • Show me the status of all agents in the swarm and list the tasks currently running with their progress.
  • Schedule a code-health check every morning at 9 and write an analysis of new Sentry alerts to a dashboard page.
  • Compile last week's task costs and token usage into a report page I can share.

Use cases

Route tasks from Slack, GitHub, GitLab, email, Linear, Jira, or the API to an agent team
Execute software development tasks in isolated Docker containers and produce pull requests
Run workflows with human approval, retries, and deterministic fan-out and join behavior
Support ongoing customer support, content generation, reporting, and code-health work with shared memory and schedules

Overview

Agent Swarm is an open-source multi-agent orchestration API and MCP server with local SQLite storage. A lead agent breaks down tasks and delegates them to worker agents running in isolated Docker environments, with shared memory, persistent identity, workflows, scheduling, and human approval gates.

Similar servers

Context7 80 · B

Upstash's official server providing up-to-date third-party library docs for AI coding assistants

★ 62.9k · Tools 2 Compare with this →
Compare with other MCP servers

Source revision 28605204d5ef Data synced 2026-10-12 Read the FMRS scoring method