Connecting OmniMem to Kiro
Kiro is AWS's AI IDE, built on VS Code, and it plans before it codes. It supports Streamable HTTP MCP servers, and its spec-driven workflow pairs rather nicely with OmniMem's project context.
Quick setup
Create or edit ~/.kiro/settings/mcp.json for global access:
{
"mcpServers": {
"omnimem": {
"type": "http",
"url": "http://127.0.0.1:8765/mcp"
}
}
}
On the desktop app, Copy MCP URL in the tray or menu bar menu gives you the address.
If OmniMem has an access token set (MCP_AUTH_TOKEN, or Access token on the Configuration page), add it as a header:
{
"mcpServers": {
"omnimem": {
"type": "http",
"url": "http://127.0.0.1:8765/mcp",
"headers": {
"Authorization": "Bearer ${OMNIMEM_TOKEN}"
}
}
}
}
Set OMNIMEM_TOKEN in your shell environment (.zshrc, .bashrc) so Kiro can expand it. The IDE uses ${VAR_NAME} syntax.
Coming from 6.x, change "type": "sse" to "type": "http" and /sse to /mcp.
Global vs project config
| File | Scope | Use when |
|---|---|---|
~/.kiro/settings/mcp.json |
All projects | OmniMem should be available everywhere (recommended) |
.kiro/settings/mcp.json (project root) |
Single project | Only this project needs OmniMem |
Kiro's MCP panel
You can also add the server through Kiro's UI:
- Open the MCP Servers panel from the sidebar or command palette
- Click Add Server
- Choose HTTP as the transport
- Enter the URL:
http://127.0.0.1:8765/mcp - Add the authorisation header if you've set a token
Using OmniMem with Kiro
Once connected, the tools are available to Kiro's agent:
- "Call the OmniMem briefing tool for this project"
- "Remember this architectural decision in OmniMem"
- "Check whether we've tried this approach before"
The agent can load project state at the start of each session and store the decisions as your specs evolve, so the reasoning behind a design doesn't evaporate when you close the window.
Verifying the connection
In Kiro's chat panel, ask:
Can you call the OmniMem health tool?
You should get back the record and vector counts per namespace, whether the embedding model is loaded, and the uptime.
Known quirks
- Silent failures: if OmniMem can't be reached, Kiro can fail to load every MCP server without saying anything. If your tools suddenly disappear, check OmniMem is running first.
- CLI vs IDE variable syntax: the IDE uses
${VAR_NAME}but the Kiro CLI expects${env:VAR_NAME}, so onemcp.jsoncan't serve both without editing. - OAuth redirects: Kiro hasn't been able to complete localhost OAuth redirects in the versions I've tried. Stick with an access token in
headers, which works fine. OAuth is for clients like claude.ai that connect from elsewhere.
Notes
- If OmniMem isn't answering,
curl http://127.0.0.1:8765/healthzshould return{"status": "ok"}, and the desktop app's tray status line says whether it's running - Check the Kiro documentation for the latest MCP options