MCP Configuration¶
Complete guide to configuring Model Context Protocol (MCP) servers in SuperQode.
Overview¶
MCP servers provide tools and resources that agents can use during SuperQode sessions and harness runs. SuperQode supports:
- Stdio Transport: Local process-based servers
- HTTP Transport: Streamable HTTP servers
- SSE Transport: Legacy Server-Sent Events servers
This page describes SuperQode consuming external MCP servers. To expose HarnessSpec workflows as an MCP server, use superqode mcp.
Configuration Location¶
MCP servers can be configured in two ways:
-
YAML Configuration (
superqode.yaml):mcp_servers: filesystem: transport: stdio command: npx args: ["-y", "@modelcontextprotocol/server-filesystem", "."] -
JSON Configuration (
.superqode/mcp.json,~/.superqode/mcp.json, or~/.config/superqode/mcp.json):{ "mcpServers": { "filesystem": { "transport": "stdio", "command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", "."] } } }
The shared resolver merges files in this order, from lowest to highest priority:
~/.config/superqode/mcp.json~/.superqode/mcp.json<working-directory>/.superqode/mcp.json- Harness
runtime.config.mcp_servers(orruntime.config.mcp) entries
A higher-priority server declaration replaces the whole earlier declaration. Disabled or invalid overrides do not silently enable an older entry. Working directories are resolved when configuration is loaded. Stdio servers default to the harness working directory. ${NAME} references are substituted in environment values and HTTP headers; commands are not shell-expanded.
Set runtime.config.mcp_config: false for inline-only harness configuration, or set it to a file path to use one explicit JSON file plus inline overrides. Relative paths resolve against the harness working directory. This resolver is shared by hosted PiPy, native MCP clients and SuperQode's harness MCP bridge; vendor SDKs can retain their own configuration paths.
ACP sessions receive enabled JSON definitions when a new session is created. Top-level YAML definitions remain available to native client sessions.
Web Fetch for ACP Agents¶
SuperQode's built-in fetch and web_fetch tools are available to the default SuperQode runtime. ACP agents such as OpenCode need fetch exposed through MCP. Add an enabled fetch server to one of the JSON MCP config locations:
{
"mcpServers": {
"fetch": {
"transport": "stdio",
"command": "uvx",
"args": ["mcp-server-fetch"],
"enabled": true
}
}
}
Restart the ACP connection after editing this file. Existing ACP sessions keep the MCP server list they received at session creation.
Transport Types¶
Stdio Transport¶
Local process-based servers (most common):
mcp_servers:
filesystem:
transport: stdio
enabled: true
auto_connect: true
command: npx
args:
- -y
- "@modelcontextprotocol/server-filesystem"
- "."
env:
NODE_ENV: production
cwd: /path/to/workspace
timeout: 30.0
Fields:
| Field | Type | Description |
|---|---|---|
transport | "stdio" | Transport type |
enabled | boolean | Whether server is enabled |
auto_connect | boolean | Auto-connect on startup |
command | string | Executable command |
args | array | Command arguments |
env | object | Environment variables |
cwd | string | Working directory |
timeout | number | Connection timeout (seconds) |
Semantic Code Search MCP¶
For AST-aware semantic repository search, install ccc and run cocoindex-code as a stdio MCP server:
uv tool install 'cocoindex-code'
mcp_servers:
cocoindex-code:
transport: stdio
enabled: true
auto_connect: true
command: ccc
args: ["mcp"]
cwd: /path/to/workspace
Run ccc init --litellm-model ollama/nomic-embed-text and ccc index in the workspace first if you want local Ollama embeddings. The MCP server exposes a semantic search tool with query, limit, offset, languages, paths, and refresh_index.
For local laptops, note that ccc mcp starts a background index task and its search tool refreshes the index by default. For repeated searches after a manual ccc index, pass refresh_index=false to keep calls lightweight. See Semantic Code Search for the native semantic_search tool, MCP tradeoffs, and local-model setup.
HTTP Transport¶
HTTP/HTTPS servers for remote or containerized MCP servers:
mcp_servers:
database:
transport: http
enabled: true
auto_connect: true
url: http://localhost:8080/mcp
headers:
Authorization: "Bearer ${MCP_DB_TOKEN}"
timeout: 30.0
sse_read_timeout: 300.0
Fields:
| Field | Type | Description |
|---|---|---|
transport | "http" | Transport type |
enabled | boolean | Whether server is enabled |
auto_connect | boolean | Auto-connect on startup |
url | string | Server URL |
headers | object | HTTP headers |
timeout | number | Request timeout (seconds) |
sse_read_timeout | number | SSE read timeout (seconds) |
SSE Transport¶
Server-Sent Events for streaming responses:
mcp_servers:
streaming:
transport: sse
enabled: true
url: http://localhost:8080/mcp/events
headers:
Authorization: "Bearer ${TOKEN}"
timeout: 5.0
sse_read_timeout: 300.0
Fields:
| Field | Type | Description |
|---|---|---|
transport | "sse" | Transport type |
enabled | boolean | Whether server is enabled |
auto_connect | boolean | Auto-connect on startup |
url | string | SSE endpoint URL |
headers | object | HTTP headers |
timeout | number | Connection timeout (seconds) |
sse_read_timeout | number | SSE read timeout (seconds) |
Common MCP Servers¶
Filesystem Server¶
Access local file system:
mcp_servers:
filesystem:
transport: stdio
enabled: true
command: npx
args:
- -y
- "@modelcontextprotocol/server-filesystem"
- "."
GitHub Server¶
Interact with GitHub repositories:
mcp_servers:
github:
transport: stdio
enabled: true
command: npx
args:
- -y
- "@modelcontextprotocol/server-github"
env:
GITHUB_TOKEN: "${GITHUB_TOKEN}"
PostgreSQL Server¶
Query PostgreSQL databases:
mcp_servers:
postgres:
transport: stdio
enabled: true
command: npx
args:
- -y
- "@modelcontextprotocol/server-postgres"
env:
POSTGRES_CONNECTION_STRING: "${POSTGRES_CONNECTION_STRING}"
Slack Server¶
Interact with Slack:
mcp_servers:
slack:
transport: stdio
enabled: true
command: npx
args:
- -y
- "@modelcontextprotocol/server-slack"
env:
SLACK_BOT_TOKEN: "${SLACK_BOT_TOKEN}"
SLACK_TEAM_ID: "${SLACK_TEAM_ID}"
Brave Search Server¶
Search the web:
mcp_servers:
brave-search:
transport: stdio
enabled: true
command: npx
args:
- -y
- "@modelcontextprotocol/server-brave-search"
env:
BRAVE_API_KEY: "${BRAVE_API_KEY}"
MCP Server Assignment¶
MCP ownership depends on the selected runtime:
| Runtime path | How MCP servers are attached |
|---|---|
Hosted PiPy, builtin HarnessSpec | Shared JSON files plus inline runtime declarations |
openai-agents HarnessSpec | SuperQode bridges discovered MCP tools as SDK function tools |
| PydanticAI | Uses its native MCP configuration path |
| Codex, Claude, and Copilot SDKs | Uses the runtime's local MCP configuration |
| ACP coding agents | SuperQode forwards enabled JSON MCP definitions during session creation |
| Google ADK | MCP tools are not bridged yet |
Project-level mcp_servers definitions are available to SuperQode-managed client sessions, but they are not automatically injected into every vendor SDK. Use the runtime-specific documentation when the runtime owns its MCP lifecycle.
Runnable Harness Example¶
The repository includes a local FastMCP documentation server and a HarnessSpec that consumes it over stdio:
superqode harness run \
--spec examples/harnesses/mcp-docs.yaml \
--prompt "Find the documentation for MCP serving"
The example server is examples/mcp/local_docs_server.py. Its tool is exposed to the harness as mcp_docs_search_docs.
Environment Variables¶
Use environment variables in MCP configuration:
mcp_servers:
github:
transport: stdio
env:
GITHUB_TOKEN: "${GITHUB_TOKEN}" # Will read from $GITHUB_TOKEN
Variable Resolution:
${VAR_NAME}: Reads from environment- Default values:
${VAR_NAME:-default}(not supported, use explicit env vars)
Enabling/Disabling Servers¶
Enable All Servers¶
mcp_servers:
filesystem:
enabled: true
github:
enabled: true
Disable Specific Server¶
mcp_servers:
filesystem:
enabled: false # Server defined but disabled
github:
enabled: true
Auto-Connect Control¶
mcp_servers:
filesystem:
enabled: true
auto_connect: true # Connect on startup
github:
enabled: true
auto_connect: false # Connect on demand
Troubleshooting¶
Server Not Starting¶
Problem: MCP server fails to start
Solution: Check command and environment:
mcp_servers:
filesystem:
command: npx
args: ["-y", "@modelcontextprotocol/server-filesystem", "."]
env:
DEBUG: "1" # Enable debug logging
Connection Timeout¶
Problem: Server connection times out
Solution: Increase timeout:
mcp_servers:
database:
transport: http
url: http://localhost:8080/mcp
timeout: 60.0 # Increase from default 30.0
Tools Not Available¶
Problem: MCP tools not showing up
Solution:
-
Verify the project configuration is valid:
superqode config validate -
Start SuperQode or the harness MCP server and watch startup output for MCP connection errors:
superqode superqode mcp --dir ./harnesses -
Verify the configured MCP server command works outside SuperQode. For stdio servers, run the configured command directly and check that it starts.
Best Practices¶
1. Use Auto-Connect Selectively¶
Enable auto-connect only for frequently used servers:
mcp_servers:
filesystem:
auto_connect: true # Always needed
github:
auto_connect: false # Only when working with GitHub
2. Organize by Purpose¶
Group related servers:
# Development servers
mcp_servers:
filesystem: {...}
git: {...}
# External services
mcp_servers:
github: {...}
slack: {...}
database: {...}
3. Use Environment Variables¶
Never hardcode secrets:
mcp_servers:
github:
env:
GITHUB_TOKEN: "${GITHUB_TOKEN}" # [CORRECT] Good
# GITHUB_TOKEN: "ghp_..." # [INCORRECT] Bad
4. Test Connections¶
Verify servers work before using:
superqode config validate
superqode mcp --dir ./harnesses
JSON Configuration Format¶
MCP servers can also be configured in JSON format:
{
"mcpServers": {
"filesystem": {
"transport": "stdio",
"enabled": true,
"autoConnect": true,
"command": "npx",
"args": [
"-y",
"@modelcontextprotocol/server-filesystem",
"."
],
"env": {
"NODE_ENV": "production"
},
"timeout": 30.0
}
}
}
Note: Hosted harness inline declarations override shared JSON entries as described above.
Advanced Configuration¶
Multiple Instances¶
Run multiple instances of the same server:
mcp_servers:
filesystem-workspace:
transport: stdio
command: npx
args: ["-y", "@modelcontextprotocol/server-filesystem", "/workspace"]
filesystem-home:
transport: stdio
command: npx
args: ["-y", "@modelcontextprotocol/server-filesystem", "~"]
Custom Working Directory¶
mcp_servers:
filesystem:
transport: stdio
command: npx
args: ["-y", "@modelcontextprotocol/server-filesystem", "."]
cwd: /path/to/project # Override working directory
Custom Headers for HTTP¶
mcp_servers:
api:
transport: http
url: https://api.example.com/mcp
headers:
Authorization: "Bearer ${API_TOKEN}"
X-API-Version: "v2"
User-Agent: "SuperQode/1.0"
Next Steps¶
- YAML Reference - Complete configuration reference
Inspect and refresh connections¶
Open :mcp panel in the TUI for live connection state, tool/resource/prompt counts, and redacted server diagnostics. Tab moves to Connect, Retry, Login, Disconnect, Cancel and Back; arrow keys select a server. Login is available for HTTP/SSE servers and may open an authorization browser. Retry uses the existing reconnect operation. Cancel stops a pending connection or login and disconnects that server.
Escape closes the panel and preserves the prompt draft. Established connections and pending transport initialization continue when the panel closes; a pending login started by the panel is cancelled. Reopen it to inspect or cancel initialization. Opening the panel alone does not connect servers or start authentication.
sq mcp list --cwd ./project
sq mcp list --connect --cwd ./project
sq mcp login SERVER
sq mcp logout SERVER
sq mcp reconnect SERVER
In the TUI, :mcp reload reloads configuration and disconnects removed or changed servers before reconnecting. :mcp login SERVER and :mcp logout SERVER use the shared OAuth store. Login may open an authorization browser when the server requires OAuth. Configuration reports omit credentials and endpoint URLs.