MCP server for coding agents
e-volv Observer serves a remote MCP server, so the agent that is about to change the code can read the same evidence you would: the log lines, the error group with its stack and source link, the incident and its explanation, and the trace. Every tool is read-only.
Endpoint and token
The server speaks the Streamable HTTP transport at https://api.e-volv.io/api/public/v1/observer/mcp. It is stateless and answers POST only.
Create a token in the app under Observer, MCP. A token is a read-only (logs:read) key for one workspace, shown once, and it expires after 30 days, 90 days or a year. Send it as Authorization: Bearer evk_.... Any server API key with logs:read works too. A token never reads another workspace, and the workspace needs e-volv Observer: a workspace without it gets 402.
A server key is refused when it arrives with an Origin header, so connect from a CLI or an editor, not from a web page.
Claude Code
claude mcp add --transport http e-volv-observer https://api.e-volv.io/api/public/v1/observer/mcp \ --header "Authorization: Bearer evk_YOUR_OBSERVER_MCP_TOKEN"
Cursor
In ~/.cursor/mcp.json, or .cursor/mcp.json in a project:
{
"mcpServers": {
"e-volv-observer": {
"url": "https://api.e-volv.io/api/public/v1/observer/mcp",
"headers": {
"Authorization": "Bearer evk_YOUR_OBSERVER_MCP_TOKEN"
}
}
}
}Codex
# ~/.codex/config.toml [mcp_servers.e-volv-observer] url = "https://api.e-volv.io/api/public/v1/observer/mcp" bearer_token_env_var = "EVOLV_OBSERVER_TOKEN" # then, in your shell: # export EVOLV_OBSERVER_TOKEN=evk_YOUR_OBSERVER_MCP_TOKEN
Jacker
Add an http tool server to a plugin manifest. Jacker asks for the token once and keeps it with its other secrets. See the manifest reference.
{
"permissions": {
"network": [
"api.e-volv.io"
],
"filesystem": "project-read",
"config": [
{
"key": "OBSERVER_TOKEN",
"label": "e-volv Observer MCP token",
"secret": true,
"required": true
}
]
},
"contributes": {
"mcpServers": [
{
"name": "e-volv-observer",
"transport": "http",
"url": "https://api.e-volv.io/api/public/v1/observer/mcp",
"headers": {
"Authorization": "Bearer ${config.OBSERVER_TOKEN}"
},
"readOnlyTools": [
"search_logs",
"list_error_groups",
"get_error_group",
"list_incidents",
"get_incident",
"get_trace"
]
}
]
}
}Tools
search_logs: Search log events with the query language.list_error_groups: Error groups with counts and culprit file:line.get_error_group: One group: source link, Code Map symbol, stack frames.list_incidents: Incidents, open first.get_incident: Failure graph and the cached explanation.get_trace: One trace as a span waterfall.
Results are compact text sized for a model’s context: at most 50 rows, long lines cut, 15 stack lines per occurrence by default, and about 12,000 characters per result. Lists page with limit and offset; occurrences page with the cursor a result names. The query language is the one on the Observer overview. get_incident returns an explanation only when one has been generated in the app; the MCP server never starts a billed explanation run.