Agent Context

Three read-only MCP servers that let your AI coding agents read each other's conversation history — so you can pick up in one tool exactly where you left off in another.

If you run more than one coding agent (Claude Code, Codex, Hermes), they normally have no idea what you did in the others. You end up re-explaining the same context several times a day. These servers fix that: each agent can read the past conversations of the others, read-only and sanitized.

Server	Reads	Source
hermes_context	Hermes sessions	state.db (SQLite)
codex_context	Codex sessions	~/.codex/sessions/**/rollout-*.jsonl
claude_code_context	Claude Code sessions	~/.claude/projects/**/*.jsonl

Safety contract

Every server is strictly read-only and sanitizes before returning anything:

• Read-only. Source transcripts/databases are opened read-only. Nothing is ever written back.
• No hidden instructions. System/developer messages, base instructions and harness bootstrap blocks (slash commands, <system-reminder>, project memory dumps) are excluded.
• No reasoning leaks. Assistant "thinking"/reasoning and encrypted chains-of-thought are dropped unless explicitly requested.
• Tools are opt-in. Tool calls/outputs are excluded by default and truncated when included.
• Secrets redacted. API keys, tokens and JWTs are replaced with [REDACTED_SECRET].
• Tolerates live writes. A half-written trailing JSON line in an active session is ignored, not fatal.

Output is structured Markdown, not raw JSON.

Requirements

• Python 3.10+
• The mcp package (installed below)

pip install -r requirements.txt

By default each server auto-detects its source under your home directory. Override with env vars if your stores live elsewhere:

• CODEX_SESSIONS_ROOT (default ~/.codex/sessions)
• CLAUDE_PROJECTS_ROOT (default ~/.claude/projects)
• HERMES_HOME (default %LOCALAPPDATA%/hermes on Windows, else ~/.hermes)

Install per agent

Use your Python interpreter (python below — use the full path to the interpreter where you installed mcp).

Claude Code

claude mcp add codex_context       --scope user -- python /path/to/agent-context-mcp/servers/codex_context_mcp.py
claude mcp add hermes_context      --scope user -- python /path/to/agent-context-mcp/servers/hermes_context_mcp.py
# (optional) read your own past Claude Code sessions:
claude mcp add claude_code_context --scope user -- python /path/to/agent-context-mcp/servers/claude_code_context_mcp.py

Codex (~/.codex/config.toml)

[mcp_servers.claude_code_context]
command = 'python'
args = ['/path/to/agent-context-mcp/servers/claude_code_context_mcp.py']

[mcp_servers.hermes_context]
command = 'python'
args = ['/path/to/agent-context-mcp/servers/hermes_context_mcp.py']

Hermes

hermes mcp add codex_context       --command python --args /path/to/agent-context-mcp/servers/codex_context_mcp.py
hermes mcp add claude_code_context --command python --args /path/to/agent-context-mcp/servers/claude_code_context_mcp.py

Restart the agent's session after adding a server so its MCP cache picks it up.

Full mesh

Register all three servers in all three agents and every agent can read every agent (its own past sessions included):

Agent	reads Hermes	reads Codex	reads Claude Code
Claude Code	✅	✅	✅
Codex	✅	✅	✅
Hermes	✅	✅	✅

Tools

• hermes_context: list_hermes_conversations, search_hermes_conversations, get_hermes_conversation_context, get_latest_hermes_conversation_context
• codex_context: codex_list_conversations, codex_search_conversations, codex_read_conversation, codex_read_latest_conversation, codex_export_context_bundle
• claude_code_context: claude_code_list_conversations, claude_code_search_conversations, claude_code_read_conversation, claude_code_read_latest_conversation, claude_code_export_context_bundle

License

MIT — see LICENSE.