AgentBridge by Graphene Lab MCP
AgentBridge Graphene Lab
The standalone AgentBridge server exposes over MCP one office agent with Word, Excel, email and a task scheduler
- 20 ★ on the repository
- Last commit
- AGPL-3.0
Broad access
Why this level
- One `agent_run` call starts an autonomous run with files, mail over SMTP and IMAP, git and a scheduler, and the run continues after the client disconnects
- The agent reads the whole documents folder, and with a cloud provider the texts go to a cloud model
- Cloud provider keys sit in `providers.json` next to the executable
Install
- Works in
Manual install
docker run -d -p 127.0.0.1:5290:5290 -v agentbridge-data:/app/PersistentData ghcr.io/graphene-lab/agentbridge:latestStarts the server with the MCP endpoint `http://localhost:5290/mcp`. The model and documents folder are set in `/setup`.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- One `agent_run` call starts an autonomous run with files, mail over SMTP and IMAP, git and a scheduler, and the run continues after the client disconnects
- The agent reads the whole documents folder, and with a cloud provider the texts go to a cloud model
- Cloud provider keys sit in `providers.json` next to the executable
This is third-party code. Review the repository files before installing.
What it does
AgentBridge runs as a separate .NET server and gives one process with three entrances: a terminal chat, an OpenAI-protocol HTTP API at `http://localhost:5290/v1` and an MCP endpoint at `http://localhost:5290/mcp`. Over MCP a client sees one tool, `agent_run`. It takes `prompt`, plus `model`, `llm_provider`, `max_iterations` and `session_id`, and starts an autonomous run with the tools the agent has enabled. The built-in set targets office work: Word documents, Excel sheets, presentations, PDF reports, web, email over SMTP and IMAP, files, git, a task scheduler, maps and FreeCAD parametric models. The agent works in a sandbox: file changes are kept with versions, and it cannot write or run code. It reads the documents folder you set in the settings. The model is chosen in `/setup`: DeepSeek, Z.ai, Gemini, Anthropic, any OpenAI-compatible service or local Ollama and ExLlamaV2.
Who it is for. Small companies and administrators who deploy an office agent for documents and mail in-house and want to drive it from Claude Desktop or VS Code.
Good fit when
- Staff need an assistant that builds a report from contracts and invoices in a shared folder and saves it as Word or Excel
- A Claude Desktop client hands off a long task, and `agent_run` carries it through even if you close the window
- Weekly scheduled work, such as an inbox digest, is set up through the built-in scheduler
- Data must not go to the cloud, so the agent is connected to a local model through Ollama
- Scripts and bots need their own OpenAI-protocol server and hit the same `localhost:5290/v1`
Not a fit when
- You need a coding agent: code-agent features are deliberately absent
- Ordinary document MCP servers are enough: AgentBridge installs a whole agent with its own model
- The AGPL v3.0 license does not fit your project
Example request
Through AgentBridge build an Excel table of end dates and amounts from the contracts folder, save it next to them and send nothing by emailMCP
- Transport
- http, stdio
- Authentication
- not required
Environment variables
AGENTBRIDGE_URL- Address of the local MCP endpoint when the bridge is run over stdio: http://localhost:5290/mcp.
Limitations
- The MCP client passes only the task text, and the work is done by AgentBridge's own agent with the model from its settings.
- Before the first task, choose a model provider and a documents folder in `/setup`.
- The archive weighs about 460 MB, but .NET does not need to be installed.
- Claude Desktop needs the bridge from `agentbridge-mcp.mcpb` and Node.js 18 or newer.
- The agent cannot write or run code, as the sandbox forbids it.
- The license is AGPL v3.0.
How to disable. Stop the container with `docker stop` or close the `agent` process. Remove the connector from the client settings: the `agentbridge-mcp.mcpb` extension in Claude Desktop or the server entry in `.vscode/mcp.json`. Model keys live in `providers.json`.
FAQ
How do I connect AgentBridge to VS Code?
Start AgentBridge, then in the command palette choose `MCP: Add Server`, type HTTP, and enter `http://localhost:5290/mcp`. A ready file is in the repository: copy `examples/vscode-mcp.json` to `.vscode/mcp.json`.
How do I connect it to Claude Desktop?
Download `agentbridge-mcp.mcpb` from the latest release page and install it through Settings, Extensions, Advanced settings, Install extension. It is a small bridge on Node.js 18 or newer that forwards the local endpoint over stdio.
What does the single agent_run tool do?
It starts an autonomous agent run from the `prompt` text. Optional parameters are `model`, `llm_provider`, `max_iterations` and `session_id` to continue a conversation. The agent uses all enabled tools and finishes the task even if the client disconnects.
How do I run AgentBridge in Docker?
The command from the repository is `docker run -d -p 127.0.0.1:5290:5290 -v agentbridge-data:/app/PersistentData ghcr.io/graphene-lab/agentbridge:latest`. The port is bound to the local address, and data lives in the `agentbridge-data` volume.
Do documents leave the company?
The documents folder stays local, only the model call goes out, and only if a cloud provider is chosen. With a local model through Ollama nothing leaves the machine. There is also an anonymization switch for names in cloud calls.
Discussion
No questions or reviews yet.
Related
MCP serversA bridge between local CLI agents: Codex, Cursor or Claude Code delegate work to Antigravity, Grok, Kimi and seven more.
An open coding agent for the terminal, browser and desktop that works with models from OpenAI and Claude to Ollama
A self-hosted AgentOS for one workspace: plan-mode chat, private knowledge base, MCP tools, skills, memory and a sandbox
A local bridge linking Claude Code and Codex in one session so the agents message each other and review each other's work
Agent Mesh links agents, containers and laptops into a private network where nodes publish tools and models over MCP
UseAgent’s MCP server sends tasks to remote agents, collects their results and records the runs in an organization’s ledger.
README in short
The page begins with a product description and a diagram of the host, the agent core and plugins, then links to installation, a demo gallery and a comparison with similar products. The MCP section is a separate block with a table of clients, followed by documents folder and provider setup, documentation and an FAQ.