intern-reader

name	intern-reader
description	Guide for querying experiment data via my-cheap-intern MCP tools. TRIGGER when: user asks about experiments, runs, metrics, training results, or logs — e.g. "今天跑了哪些实验", "最新结果怎么样", "对比这两个run", "训练炸了吗", "show today's runs", "compare experiments", "is it overfitting", "what crashed". Also trigger when you need to look up run IDs, check training progress, or investigate failures.

my-cheap-intern Experiment Reader

Query experiment data through 6 MCP tools.

MCP Connection

• Server name: my-cheap-intern (configured by intern-cli init)
• Tool name pattern (Claude Code): mcp__my-cheap-intern__{tool_name}
  • Example: mcp__my-cheap-intern__list_projects, mcp__my-cheap-intern__search_runs
• Pre-check: If mcp__my-cheap-intern__list_projects is not in your available tools, the server is not connected. Tell the user to start the server (intern-server launch) and run intern-cli init in their project to configure .mcp.json.

MCP Tools Reference

list_projects

List all projects with run count and last active time.

No parameters.

search_runs

Search runs within a project.

Param	Type	Required	Description
project	string	yes	Project name
query	string	no	Fuzzy search across run name, tags, config values
tags	string[]	no	Exact tag match (AND logic)
status	enum	no	"running" / "finished" / "crashed"
started_after	string	no	ISO datetime
started_before	string	no	ISO datetime

get_run_summary

Pre-computed summary with trends, anomalies, log stats.

Param	Type	Required	Description
run_ids	string[]	yes	One or more run IDs

Returns per run: config, status, duration, metrics (last/min/max/mean + trend direction), log_summary.

compare_runs

Side-by-side comparison: config diff + metrics ranking.

Param	Type	Required	Description
run_ids	string[]	yes	Two or more run IDs
metric_keys	string[]	no	Defaults to all shared metric keys

Returns: config_diff (only differing keys), metrics_table, ranking.

get_metric_series

Raw time series data for one metric.

Param	Type	Required	Description
run_id	string	yes	Single run ID
key	string	yes	Metric key name
start_step	int	no	Filter by step range
end_step	int	no	Filter by step range
downsample	int	no	Downsample to N points

get_logs

Text log entries with filtering.

Param	Type	Required	Description
run_id	string	yes	Single run ID
level	enum	no	"info" / "warning" / "error"
keyword	string	no	Substring search
limit	int	no	Default 100

Common Query Patterns

"Show me today's experiments"

1. list_projects to find the project name
2. search_runs(project=..., started_after="2026-04-01T00:00:00") to get today's runs
3. get_run_summary(run_ids=[...]) for overview with trends

"Compare two runs"

1. compare_runs(run_ids=[id1, id2]) for config diff and metrics ranking
2. If deeper look needed: get_metric_series for specific curves

"Is this run overfitting?"

1. get_run_summary(run_ids=[id]) — check if train_loss trend is "decreasing" but val_loss trend is "increasing"
2. get_metric_series(run_id=id, key="train_loss") and get_metric_series(run_id=id, key="val_loss") for the raw curves

"What went wrong with this crashed run?"

1. get_run_summary(run_ids=[id]) — check status and duration
2. get_logs(run_id=id, level="error") — find error messages
3. get_logs(run_id=id, keyword="nan") — check for NaN issues

"Find the best run for a given metric"

1. search_runs(project=..., status="finished") — get all finished runs
2. compare_runs(run_ids=[...], metric_keys=["val_acc"]) — ranking will show the best