agentcontextDownload
OPEN SOURCE · SELF-HOSTED · MCP

Different agents.
Same understanding.

Shared context for MCP-compatible coding agents.
Share progress, decisions, and handoffs across cloud-backed agents and compatible local-model workflows.

Download Agent Context Get started
$npm install -g ./agent-context-mcp-0.4.0.tgz

After downloading, run this in the folder containing the .tgz file.

Node.js 24+Runs on your machineMIT license
ONE PROJECT. A SHARED THREAD.↙ TRY THE HANDOFF
launchpad / shared contextINTERACTIVE EXAMPLE
✳
ClaudeRecords the work
handoff
PROJECT CONTEXT

Authentication route complete

Added session validation and covered the failure cases.

↳src/auth.ts
↗Next: connect the login screen.
Persistent contextSample project data
↳The next agent starts with context.
YOUR TOOLS.
CONNECTED THROUGH MCP.
✳Claude Code⌘Codex◩Cursor+ other MCP clients

YOUR MODEL. YOUR MACHINE. SHARED CONTEXT.

Connect through MCP.
Keep your choice of tools.

Agent Context works through your agent’s MCP integration. Choose a compatible client, share one project record, and keep building.

⌘

Coding agents

Use Claude Code, Codex, Cursor, VS Code/Copilot, Gemini CLI, or another client with MCP tools over stdio or Streamable HTTP. Enable the tools in your client.

Choose a client
◫

Windows, macOS, and Linux

Designed for desktop systems with Node.js 24+. Local agents share a database path; remote agents connect to one reachable HTTP server.

System requirements
⬡

Local-model workflows

Use a tool-capable local model with an MCP-capable agent or host. Agent Context does not require a cloud model or a model API key.

Local model setup

Compatibility depends on the client, model, and tool configuration. Automated protocol checks passed locally; every client and operating-system combination has not been tested.

DIFFERENT COMPUTERS. ONE PROJECT RECORD.

Your team connects.
The context stays together.

Run one shared server. Teammates keep their own coding agents and models, and connect through the network method that fits your team.

Same Wi-Fi / LAN

Connect to the host’s private IP on your office or home network. Configure the host firewall and exact allowlists.

LAN setup

Private VPN

Use a private encrypted network for teammates working from different locations, with a reachable private endpoint.

VPN setup

SSH tunnel

Forward a local port to the shared server over SSH. Keep the context service bound to loopback on the host.

SSH setup

Shared HTTPS server

Run on persistent infrastructure with a domain and TLS reverse proxy for an always-on team endpoint.

HTTPS setup

Use the host’s token, the same project ID, and a unique agent ID per session. Plain LAN HTTP is unencrypted; prefer VPN, SSH, or HTTPS for routine team access. All token holders share access to the instance.

THE PROJECT CONTINUES. THE CONTEXT SHOULD TOO.

Your code has history.
Give your agents some.

A session ends. A different agent takes over. Keep the important context available before the next edit.

▤

A shared project memory.

Architecture, progress, file references, and the decisions behind them. Stored durably and retrieved through MCP.

get_context(project_id)
↗

A handoff with a next step.

Record what changed, what was tested, what is blocked, and exactly where the next agent should pick up.

record_update(kind: "handoff")
◷

Work with a clear owner.

Claim tasks with expiring leases and version checks. See who holds a task before another agent starts it.

claim_task(expected_version)

THREE MOMENTS. ONE CONTINUOUS BUILD.

The handoff
becomes the starting point.

Agents read and write the same project record. You choose the workflow; the server keeps the context available.

  1. 01

    Read before editing.

    Load the project context and inspect current tasks.

  2. 02

    Record the meaningful changes.

    Keep decisions, file references, and test results with the work.

  3. 03

    Leave a useful next step.

    Write a handoff that the next session can act on.

Set up the agent workflow
HANDOFF.MDILLUSTRATIVE PROJECT RECORD

01# Ready for the next session

02

03completed: added login route

04files: src/auth.ts

05decision: validate sessions server-side

06verification: route tests passed

07next: connect the login screen

next agent →get_context("launchpad")context retrieved ✓

Saved records are agent reports. Verify the actual code and tests.

SEE THE THREAD. NOT JUST THE LATEST MESSAGE.

A workspace for
the whole build.

Projects, contributing agents, task claims, searchable context, and activity. All in the React dashboard that ships with your local server.

Workspace / OverviewLOCAL INSTANCE
YOUR WORKSPACE, IN SYNC.

Everything that comes next.

+ New project
Projects2Sample project data
Open tasks3Sample project data
Recent updates8Sample project data
Contributing agents2Sample project data

Your projects

LPLaunchpadlaunchpad / shared context↗

DKDeveloper kitdeveloper-kit / shared context↗

Latest handoff

↗
Login route complete

Next: connect the login screen.

Claude · launchpad · handoff
React dashboard included. No separate frontend server.Run your own workspace

FROM DOWNLOAD TO FIRST HANDOFF

Install. Connect.
Keep building.

One local process serves the dashboard and MCP endpoint. Bring your agents. Keep your project context on your machine.

01 / INSTALL

Start with the package.

Download the release, then open a terminal in the folder containing the .tgz file (usually Downloads).

TERMINAL · PACKAGE FOLDER
npm install -g ./agent-context-mcp-0.4.0.tgz
Download v0.4.0 Installing from the source repository? Node.js 24+ · MIT licensed · npm registry release pending
02 / RUN

Your workspace, locally.

Start the service, then retrieve your instance token in another terminal.

TERMINAL
agent-context-mcp
agent-context-mcp token
localhost:8765/dashboardPaste the token into the dashboard to connect.
03 / CONNECT

Give your agents a shared memory.

Point a client at the MCP endpoint. Use the same project ID across agents.

TERMINAL
export AGENT_CONTEXT_TOKEN="$(agent-context-mcp token)"
codex mcp add agent-context --url http://127.0.0.1:8765/mcp --bearer-token-env-var AGENT_CONTEXT_TOKEN
Claude, Codex & other clients

YOURS TO RUN. OPEN TO IMPROVE.

Read it. Run it.
Make it yours.

Open source under the MIT license. Keep your data on your infrastructure, inspect how it works, and help shape what comes next.

Explore on GitHub Contribute
DEPLOYMENTSelf-hosted

One trusted user or team per instance.

STORAGELocal SQLite

History stays outside the installation directory.

PROTOCOLMCP

10 tools, plus a project resource and resume prompt.

GOOD QUESTIONS.

A little more
context.

Will it work with any AI coding agent?

It works with clients that support MCP tools over stdio or Streamable HTTP. An AI model or chat app without an MCP integration cannot connect by itself. See the compatibility guide for requirements and verification limits.

Can I use a local model?

Yes, through an MCP-capable agent or host using a model that can call tools reliably. A local inference runtime such as Ollama or llama.cpp needs an MCP-capable host or integration in the workflow. Model, template, and client support determine whether tool calls work. See local model setup.

Can agents on different computers share context?

Yes. Connect every client to one running MCP server over Streamable HTTP, using its reachable URL and instance token. Connect over the same Wi-Fi/LAN, a private VPN, an SSH tunnel, or an HTTPS server. Follow the team connection guide. The public website serves documentation and downloads; you run the MCP service separately.

Does it automatically read my agents' conversations?

Agents use MCP tools to read and record engineering context. Add the workflow to the instruction file your client reads, such as AGENTS.md, CLAUDE.md, GEMINI.md, or editor rules so each session reads the shared record and leaves a handoff. Initial project knowledge comes from an agent inspecting your repository and recording it.

Why does the dashboard ask for an access token?

The token protects the local HTTP service. Run agent-context-mcp token in another terminal and paste its output into the dashboard. If you use a custom database, pass the same --db path to both commands.

Can I use it with my existing project?

Yes. Run your agents from the existing repository and connect them to this server. Give every agent the same project ID. The server stores context in a separate SQLite file and does not edit the repository itself.

Can multiple agents edit the same project?

Yes. Task claims and version checks coordinate ownership, and expired leases can be reclaimed. Claims apply to tasks; agents should still check file overlap and use separate Git worktrees when editing concurrently.

Is there a hosted account or software subscription?

This release is self-hosted and MIT licensed. There is no product subscription or hosted signup. You operate the server and infrastructure, and any AI tools you connect retain their own accounts and costs.

THE NEXT AGENT IS READY.

Give it a good
place to start.

Download v0.4.0

Node.js 24+ · Local-first · MIT licensed