CacheBash MCP server
Commit Nov 13, 2026Enables multi-agent coordination through task dispatching, relay messaging, persistent memory, and human-in-the-loop mobile approvals.
Also known as rezzedai/cachebash
- Stars
- 0
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- 2.0.0
- Last commit
- Nov 13, 2026
- License
- MIT
What is CacheBash MCP?
CacheBash enables multi-agent coordination through task dispatching, relay messaging, persistent memory, and human-in-the-loop mobile approvals.
Summary generated from the project's README
About CacheBash MCP server
CacheBash provides a coordination layer for multi-agent systems using the Model Context Protocol. It allows different agents to communicate, hand off tasks, and maintain persistent memory across sessions. Developers can deploy the server on Cloud Run or any Node.js host, utilizing Google Cloud Firestore for isolated data storage. The system supports various clients including Claude Code, Cursor, VS Code with Copilot, and Gemini CLI via Streamable HTTP.
The platform features 108 tools organized into 21 domains, such as dispatch, relay, and state. It manages the full task lifecycle through tools for creating, claiming, and completing work, while ensuring two agents cannot claim the same task via Firestore transactions. Security is handled through capability-scoped API keys that filter the available tool list. A unique aspect of this server is the human-in-the-loop capability, which uses a React Native mobile app to allow users to approve requests or answer questions directly from a phone.
What you can do with it
- 1Coordinate multi-agent task handoffsAssign a specific piece of work to another agent and track its progress through the lifecycle.
- 2Implement human-in-the-loop approvalsPause an autonomous agent workflow to wait for a user to approve a decision via a mobile device.
- 3Maintain long-term agent memoryEnsure an agent remembers learned patterns and context history across different sessions.
- 4Enforce granular tool permissionsRestrict an agent's access to specific functions like reading or writing messages using scoped keys.
Written from the project's README. Every use case is checked against a quote from it.
How to install CacheBash MCP
Formats verified 10 Nov 2026Add to CursorAdd to GooseInstall CacheBash MCP in Claude Code
$ claude mcp add --transport http cachebash https://cachebash-mcp-922749444863.us-central1.run.app/v1/mcpInstall CacheBash MCP in Cursor
~/.cursor/mcp.json
{
"mcpServers": {
"cachebash": {
"url": "https://cachebash-mcp-922749444863.us-central1.run.app/v1/mcp"
}
}
}Install CacheBash MCP in VS Code
.vscode/mcp.json
{
"servers": {
"cachebash": {
"type": "http",
"url": "https://cachebash-mcp-922749444863.us-central1.run.app/v1/mcp"
}
}
}Install CacheBash MCP in VS Code CLI
$ code --add-mcp '{"name":"cachebash","type":"http","url":"https://cachebash-mcp-922749444863.us-central1.run.app/v1/mcp"}'Install CacheBash MCP in Codex
$ codex mcp add cachebash --url https://cachebash-mcp-922749444863.us-central1.run.app/v1/mcpInstall CacheBash MCP in Codex config.toml
~/.codex/config.toml
[mcp_servers.cachebash] url = "https://cachebash-mcp-922749444863.us-central1.run.app/v1/mcp"
Install CacheBash MCP in Gemini CLI
$ gemini mcp add -s user --transport http cachebash https://cachebash-mcp-922749444863.us-central1.run.app/v1/mcpInstall CacheBash MCP in Gemini settings.json
~/.gemini/settings.json
{
"mcpServers": {
"cachebash": {
"url": "https://cachebash-mcp-922749444863.us-central1.run.app/v1/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 reachablerezzedai/cachebash is public and reachablegithub.com/rezzedai/cachebash ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-13github.com/rezzedai/cachebash ↗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 handshakeHTTPS endpoint answered 401 (sign-in required)Pass
README
View on GitHubCacheBash
CacheBash is an open-source MCP server for multi-agent coordination. It is written in TypeScript and works with any model and any MCP client. Agents use it to hand work to each other, message each other, report their state and ask a human for decisions. The human answers from a phone.
Architecture at a glance
- 108 MCP tools in 21 domains. The default deployment profile serves 107 of them. A smaller "lite" profile for tenant deployments serves 96.
- Task dispatch. Agents create, claim and complete tasks. Claims run inside a Firestore transaction, so two agents cannot claim the same task.
- Relay messaging. Direct messages, group multicast and message history. Messages that cannot be delivered go to a dead-letter queue for inspection.
- Human in the loop. Agents send questions and approval requests. The mobile app shows them, and the agent continues when the human answers.
- Capability-scoped keys. Each API key carries capabilities such as
dispatch.readorrelay.write. The tool list a client sees is filtered to that key's capabilities, and every call is checked again. - Policy checks before dispatch. The dispatch pre-flight evaluates the configured policies before a task is created.
- Per-task telemetry. Task completion records the status, model, provider, input and output tokens, cost, and a failure class.
- Persistent agent memory. Each program keeps state, learned patterns and memories across sessions.
How it runs:
- Server: TypeScript on Cloud Run, or any Node.js host.
- Storage: Google Cloud Firestore, with data isolated per user.
- Auth: Firebase Auth (GitHub and Google sign-in) plus per-program API keys.
- Mobile: React Native and Expo, for iOS and Android.
- Transport: MCP Streamable HTTP with Bearer token auth.
┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│ Claude Code │ │ Cursor │ │ VS Code │
└──────┬──────┘ └──────┬──────┘ └──────┬──────┘
└────────────────┼────────────────┘
│ MCP (Streamable HTTP)
▼
┌───────────────────┐
│ CacheBash server │
│ (Cloud Run) │
└─────────┬─────────┘
┌───────┴────────┐
▼ ▼
┌────────────┐ ┌────────────┐
│ Firestore │ │ Mobile app │
└────────────┘ └────────────┘
Supported clients
| Client | Status |
|---|---|
| Claude Code | Supported |
| Cursor | Supported |
| VS Code + Copilot | Supported |
| Gemini CLI | Supported |
| Any MCP-compatible client | Supported |
CacheBash uses the standard MCP protocol. No orchestration logic depends on one vendor's SDK.
Tool domains
Counts come from the tool definitions in services/mcp-server/src/tools/. Most tools are named <domain>_<action>.
| Domain | Tools | What it does |
|---|---|---|
| dispatch | 24 | Task lifecycle: create, claim, complete, approve, retry, reassign, escalate, lineage, telemetry |
| gsp | 9 | Grid State Protocol: bootstrap, read, write, search, diff, propose and resolve shared state |
| pulse | 9 | Sessions and fleet health: create, update, list, pause, resume, timeline, snapshots |
| metrics | 8 | Cost summary and forecast, operational and comms metrics, SLA compliance, rate-limit events |
| state | 8 | Per-program state and memory: store, recall, reinforce, delete, context history |
| relay | 7 | Messages, directives, groups, history and dead letters |
| policy | 6 | Create, read, update, delete, list and check policies |
| schedule | 5 | Recurring and one-off scheduled work |
| sprint | 5 | Sprints with stories, dependencies and wave-based execution |
| keys | 4 | API key create, list, revoke and rotate |
| webhook | 4 | Register, list and delete webhooks, and read delivery logs |
| clu | 3 | Ingest, analyze and report on transcripts |
| signal | 3 | Push alerts and questions to the mobile app, and read the answers |
| audit | 2 | Audit log and acknowledgement compliance |
| dream | 2 | Overnight autonomous runs with budget caps |
| pattern | 2 | Consolidate learned patterns across programs |
| programs | 2 | List and update registered programs |
| trace | 2 | Query request traces |
| admin | 1 | Account administration |
| feedback | 1 | Submit feedback |
| request_help | 1 | Cross-grid help requests (lite profile only) |
| Total | 108 |
API reference: docs/api-reference.md and docs.rezzed.ai
Quick start
Hosted
# 1. Install and authenticate
npx cachebash init
# 2. Verify connection
cachebash ping
# 3. Open the mobile app. Your agent appears.
npx cachebash init signs you in through the browser and writes the MCP config for Claude Code, Cursor or VS Code.
Self-hosted
# 1. Clone the repo
git clone https://github.com/rezzedai/cachebash.git
cd cachebash
# 2. Install dependencies
cd services/mcp-server && npm install
# 3. Set up Firebase
# Create a Firebase project and enable Firestore.
# Download a service account key and set GOOGLE_APPLICATION_CREDENTIALS.
# 4. Configure environment
cp .env.example .env
# Edit .env with your Firebase project details
# 5. Start the server
npm start
# 6. Configure your MCP client
# Add to your Claude Code / Cursor MCP config:
{
"mcpServers": {
"cachebash": {
"url": "http://localhost:3000/v1/mcp",
"headers": {
"Authorization": "Bearer YOUR_API_KEY"
}
}
}
}
Set CACHEBASH_PROFILE=lite to serve the smaller tool set. The profile logic is in services/mcp-server/src/tools/index.ts. Deployment notes: docs/deployment.md.
Project structure
cachebash/
├── apps/
│ └── mobile/ # Mobile app (React Native + Expo)
├── services/
│ ├── mcp-server/ # MCP server (TypeScript, Cloud Run)
│ └── functions/ # Cloud Functions (auth, notifications, cleanup)
├── packages/
│ ├── cli/ # CLI tool (npx cachebash)
│ └── types/ # Shared TypeScript types
├── infra/ # Firebase rules and indexes
├── docs/ # Documentation
├── turbo.json # Turborepo build orchestration
└── package.json # npm workspaces root
More detail: docs/architecture.md
Contributing
The CacheBash MCP server is MIT licensed. Contributions are welcome.
git clone https://github.com/rezzedai/cachebash.git
cd cachebash
npm install # installs all workspaces
npm run dev
See CONTRIBUTING.md for guidelines.
Links
License
MIT. See LICENSE.
Built by Rezzed.ai
CacheBash MCP questions
How do I add CacheBash MCP to Claude Code?
Run claude mcp add --transport http cachebash https://cachebash-mcp-922749444863.us-central1.run.app/v1/mcp in your terminal. Add --scope user to make it available in every project.
How do I add CacheBash MCP to Cursor?
Open ~/.cursor/mcp.json and add the server under mcpServers, using the Cursor config shown in the install section above.
How do agents interact with humans?
Agents can send questions and approval requests which are displayed on a mobile app, allowing the human to provide answers so the agent can continue.
Can multiple agents claim the same task?
No, claims are processed within a Firestore transaction to ensure that two agents cannot claim the same task simultaneously.
Does it support persistent memory?
Yes, each program maintains state, learned patterns, and memories across different sessions.