AgentStack MCP server
Commit Nov 15, 2026Provides a unified backend for AI agents to execute batched actions including hosting, storage, authentication, and payments through a single tool.
Also known as agentstacktech/AgentStack, agentstack
- Stars
- 2
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- 1.0.0
- Last commit
- Nov 15, 2026
- License
- —
What is AgentStack MCP?
AgentStack provides a unified backend for AI agents to execute batched actions including hosting, storage, authentication, and payments through a single tool.
Summary generated from the project's README
About AgentStack MCP server
AgentStack provides a unified backend designed to help AI agents manage complex, multi-step tasks through a structured execution loop. Instead of forcing an agent to search through a massive tool catalog for every turn, this server uses a Work Graph approach to provide the specific next step required for a project. This system allows agents to claim and execute actions, manage project context, and handle live data updates through a safe circle process involving sandboxing and review.
Developers can integrate this via the agentstack.execute tool to maintain autonomous loops. The server supports a wide range of capabilities including hosting, storage, authentication, and payments. It is built to work with various environments through specific plugins for Cursor, Claude, ChatGPT, and VS Code. By using the agents.work_next and packet.next_action commands, an agent can move from one task to the next without re-reading the entire tool menu, making it suitable for long-running autonomous workflows.
What you can do with it
- 1Maintain autonomous task loopsAn AI agent can follow a continuous sequence of actions by requesting the next step in a project queue.
- 2Deploy live content safelyA developer can use the safe circle to ensure agent-generated copy passes checks before being published to customers.
- 3Reduce tool discovery overheadAn agent can avoid the inefficiency of searching a large catalog by receiving direct instructions for the next action.
- 4Manage complex project statesAn agent can track whether it is currently working, waiting for a dependency, or has no tasks left.
Written from the project's README. Every use case is checked against a quote from it.
How to install AgentStack MCP
Formats verified 10 Nov 2026Add to CursorAdd to GooseInstall AgentStack MCP in Claude Code
$ claude mcp add --transport http agentstack https://agentstack.tech/mcpInstall AgentStack MCP in Cursor
~/.cursor/mcp.json
{
"mcpServers": {
"agentstack": {
"url": "https://agentstack.tech/mcp"
}
}
}Install AgentStack MCP in VS Code
.vscode/mcp.json
{
"servers": {
"agentstack": {
"type": "http",
"url": "https://agentstack.tech/mcp"
}
}
}Install AgentStack MCP in VS Code CLI
$ code --add-mcp '{"name":"agentstack","type":"http","url":"https://agentstack.tech/mcp"}'Install AgentStack MCP in Codex
$ codex mcp add agentstack --url https://agentstack.tech/mcpInstall AgentStack MCP in Codex config.toml
~/.codex/config.toml
[mcp_servers.agentstack] url = "https://agentstack.tech/mcp"
Install AgentStack MCP in Gemini CLI
$ gemini mcp add -s user --transport http agentstack https://agentstack.tech/mcpInstall AgentStack MCP in Gemini settings.json
~/.gemini/settings.json
{
"mcpServers": {
"agentstack": {
"url": "https://agentstack.tech/mcp",
"type": "http"
}
}
}Automated checks
Ran Nov 15, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachableagentstacktech/agentstack is public and reachablegithub.com/agentstacktech/agentstack ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)No licenseFinding
- Commit or release in the last yearLast commit or release 2026-11-10github.com/agentstacktech/agentstack ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeinitialize answered with protocol 2025-06-18 in 242 msPass
README
View on GitHubAgentStack
After one bind, the agent does not shop the catalog. Work Graph hands it the next step. Live edits stay on the safe circle: write a copy, pass the checks, then publish.
One MCP tool, agentstack.execute. The server holds the plan. Each turn is one action. Idle is a valid answer. Blocked names the unblock.
Story: agentstack.tech/work-graph · Contract: docs/plugins/WORK_GRAPH.md · Safe circle: agentstack.tech/mcp-docs#safe-cycle
The loop
session → context.project_id → agents.work_next → packet.next_action → claim / execute → work_next
Recipe mcp_work_loop_v1. Prompt agentstack_closed_loop_autonomy.
flowchart LR
S[Session] --> B[Bind project]
B --> W[work_next]
W --> N[next_action]
N --> X[Claim and execute]
X --> W
When that step changes live data, the same project runs the safe circle:
connect → sandbox → review → promote
The agent writes a generation. Customers stay on the previous copy until the checks pass. The project’s own rule publishes it: at once, or in stages.
| Who | Without the loop | With Work Graph |
|---|---|---|
| Founder with open loops | Every new chat re-reads the tool menu | The outcome is written once. The next chat takes the head of the queue |
| A person with more tasks than hours | The agent invents a tool and spends the window choosing | The list says ready, blocked, or idle |
| A business that needs the work finished | Thrash between catalog, guess, and retry | Claim, execute, verify, on the same project as the site and the payments |
| Someone wiring an agent | Paste the catalog into the prompt | agents.work_next, then packet.next_action |
Measured on 2026-10-07
Local registry build. Not a production traffic sample. Not a provider tokenizer. Bytes ÷ 4 is an English size estimate.
| Payload | UTF-8 bytes | Rough tokens |
|---|---|---|
Example next_action plus two alternatives |
250 | ~63 |
discovery.search, first 20 compact rows |
8,889 | ~2,222 |
| Handshake contract JSON, once | 11,297 | ~2,824 |
| Compact search, 50 rows (HTTP max) | 20,852 | ~5,213 |
| Full registry catalog JSON (795 actions in that build) | 1,680,071 | ~420,018 |
The full catalog is 6,720× that example packet. A 20-row search page is 35.6× it. The public catalog you should quote is 714 actions — live GET /mcp/actions/summary. Do not add the byte ratio and the operator targets into one savings claim.
| Signal | Target |
|---|---|
Autonomous loops that follow packet.next_action |
≥ 90% |
| Catalog shopping while a packet is already present | ≤ 5% |
Those are steady-production goals, not a live counter. Method and scenarios: docs/plugins/WORK_GRAPH.md.
Docs
This repository is the public documentation (English, user- and integrator-facing): web product, MCP, plugins, REST, RAG, sandboxes, and examples.
If you use the website: docs/USER_FEATURES_GUIDE.md.
If you build a product: docs/BUILD_YOUR_PRODUCT.md — @agentstack/sdk, optional genetic-ai-starter.
If you run agents on a large codebase: docs/genetic-system/.
If you integrate: docs/MCP_AND_ECOSYSTEM.md — start with the loop, then the catalog (714 actions · docs/MCP_SCALE.md).
REST / OpenAPI: docs/OPENAPI.md — Swagger · openapi.json.
npm SDK: @agentstack/sdk · docs/sdk/
Full index: docs/README.md · What's new: docs/WHATS_NEW.md
Open source kit
| Repo | Branch | Notes |
|---|---|---|
| genetic-ai-starter | main |
Map-first install: npx @agentstack/genetic-ai-starter init. SoT lives in this repo under genetic-ai-starter/ when you have the monorepo. |
Plugins
| Plugin | Notes |
|---|---|
| cursor-plugin | v0.4.18, gen3 — rules, skills, commands, agents, hooks; Plugin MCP Connect and OAuth 2.1 device code |
| claude-plugin | Claude Desktop / API installers |
| gpt-plugin | ChatGPT / OpenAI ecosystem |
| vscode-plugin | VS Code marketplace distribution |
Product
- Site: agentstack.tech
- Work Graph: agentstack.tech/work-graph
- GitHub org: github.com/agentstacktech
AgentStack MCP questions
How do I add AgentStack MCP to Claude Code?
Run claude mcp add --transport http agentstack https://agentstack.tech/mcp in your terminal. Add --scope user to make it available in every project.
How do I add AgentStack MCP to Cursor?
Open ~/.cursor/mcp.json and add the server under mcpServers, using the Cursor config shown in the install section above.
How does the agent interact with AgentStack?
The agent uses one MCP tool, agentstack.execute, which allows it to follow a plan where each turn is one action.