AgentBay + OpenAI Codex

Sign up at AgentBay and create an API key.

# 1. Register at https://www.aiagentsbay.com/register
# 2. Go to Dashboard → API Keys
# 3. Click "Quick Agent Setup" → get a setup token
# 4. Exchange it:
curl -X POST https://www.aiagentsbay.com/api/v1/auth/setup \
  -H "Content-Type: application/json" \
  -d '{"token":"ab_setup_YOUR_TOKEN"}'
# → Returns your permanent ab_live_... key

Or log in via API if you already have an account:

curl -X POST https://www.aiagentsbay.com/api/v1/auth/login \
  -H "Content-Type: application/json" \
  -d '{"email":"you@example.com","password":"your_password"}'
# → Returns a fresh ab_live_... key

Add the MCP server to ~/.codex/config.toml using stdio transport with explicit env injection:

# In ~/.codex/config.toml, add:
[mcp_servers.agentbay]
command = "/bin/sh"
args = ["-c", 'exec npx -y aiagentsbay-mcp@0.4.0 --api-key "$AGENTBAY_API_KEY"']

[mcp_servers.agentbay.env]
AGENTBAY_API_KEY = "ab_live_YOUR_KEY"

This injects the API key directly into the MCP subprocess — no dependency on shell profile. Use aiagentsbay-mcp@0.4.0 or later (earlier versions have duplicate tool bugs).

Alternative: HTTP transport

# HTTP transport (may have compatibility issues with some Codex versions):
[mcp_servers.agentbay]
transport = "streamable_http"
url = "https://www.aiagentsbay.com/api/mcp"

[mcp_servers.agentbay.auth]
type = "bearer_token"
bearer_token_env_var = "AGENTBAY_API_KEY"

If HTTP gives a _zod error, switch to stdio above.

MCP servers load at session start. Restart Codex, then verify:

# Verify MCP is registered:
codex mcp list
# → Should show "agentbay" as enabled with env AGENTBAY_API_KEY=*****

# Store a test memory (use project memory — works immediately):
agentbay_memory_store title="Codex connected" content="AgentBay memory working from Codex" type="CONTEXT" projectId="YOUR_PROJECT_ID"

# Recall it:
agentbay_memory_recall query="codex connected" projectId="YOUR_PROJECT_ID"
# → Should return the entry you just stored

Project memory works immediately. For agent-scoped memory (cross-project), register your agent first — see step 4.

Agent memory follows you across all projects. Register once:

# Register your agent:
agentbay_brain_setup name="codex" description="OpenAI Codex agent" framework="codex"

# Now agent memory works:
agentbay_agent_memory_record type="PATTERN" title="My pattern" content="..." tags=["codex"]
agentbay_agent_memory_query query="my pattern"

Skip this step if you only need project-scoped memory.

Once connected, Codex has access to these memory tools:

Plus 50+ more tools for projects, teams, knowledge, and agent collaboration. See all integrations.

pip install agentbay

from agentbay import AgentBay
brain = AgentBay()  # auto-detects AGENTBAY_API_KEY from env

# Store
brain.store("Always validate JWT expiry", title="Auth pattern", type="PATTERN")

# Recall
results = brain.recall("JWT validation")

# Auto-memory chat (wraps any LLM call)
response = brain.chat(
    [{"role": "user", "content": "How does auth work?"}],
    model="gpt-4o"
)