agent-context-mcp

One privacy-first MCP server for developers who work across AI coding agents.

Move between Claude Code, Codex, and Hermes without re-explaining what happened. agent-context-mcp reads the useful parts of your local agent history, sanitizes them, and returns structured Markdown through one MCP interface.

Local-first. Read-only. Sanitized before output.

Why this exists

When you switch AI coding tools, each one starts without the decisions, failures, and next steps from the others. This server gives a selected client a safe, read-only view of local context from:

• Claude Code

• Codex

• Hermes

It also discovers local AI harnesses, MCP configurations, and CLI runtimes — and can generate dry-run proposals for skills or MCP migration.

Related MCP server: Colab MCP

Your sessions stay local

agent-context-mcp is not a session-sync product.

• No hosted service by default

• No telemetry or analytics

• No background uploads

• No writes to source session stores

• No secrets, raw configuration values, or auth headers returned

• No remote network connection unless you explicitly enable and deploy one

Before output, source adapters exclude system/developer instructions, bootstrap dumps, encrypted/internal reasoning, and secret-shaped values. Tool output is opt-in and bounded. Active JSONL files with a partial trailing line are tolerated.

Never commit your agent data. This repository ignores session stores, JSONL files, local databases, credentials, and .env files. Its tests check for personal Windows paths and project-specific markers.

Install

From source (current)

git clone https://github.com/mosafly/agent-context-mcp.git
cd agent-context-mcp
python -m pip install .

Requires Python 3.10+.

Claude Code

claude mcp add --scope user agent_context -- agent-context-mcp

Codex CLI

codex mcp add agent_context -- agent-context-mcp

Hermes

hermes mcp add agent_context --command agent-context-mcp

Claude Desktop

Add this to claude_desktop_config.json and restart Claude Desktop:

{
  "mcpServers": {
    "agent_context": {
      "command": "agent-context-mcp",
      "args": []
    }
  }
}

Start a new client session after installing so the MCP tool cache refreshes.

One unified interface

All conversation tools take a source:

hermes | codex | claude_code

Migration proposals never edit configuration, copy skills, install packages, or transfer credentials.

Supported clients

Cloud clients cannot access your localhost or execute a local stdio command. Do not expose this server directly to the public internet.

Optional local HTTP mode

The default transport is stdio. A local Streamable HTTP server is available only after an explicit opt-in:

AGENT_CONTEXT_ALLOW_HTTP=1 agent-context-mcp \
  --transport streamable-http \
  --host 127.0.0.1 \
  --port 8000

For Cowork, claude.ai, ChatGPT, or any cloud client, place an authenticated HTTPS gateway in front of the service. Require OAuth or equivalent authentication, rate limits, and a deliberate deployment decision. Never publish a direct session reader without those controls.

Configuration

The bundled adapters use local defaults for common session locations. Set these environment variables only when your stores live elsewhere:

CODEX_SESSIONS_ROOT
CLAUDE_PROJECTS_ROOT
HERMES_HOME

Paths are read locally; their contents are never uploaded by this package.

Development

python -m pip install -e .
python -m unittest discover -s tests
python -m py_compile src/agent_context_mcp/*.py src/agent_context_mcp/adapters/*.py

License

MIT. See LICENSE.