exp-agent
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
| EXPAGENT_RUNS_DIR | No | Path to the experiment runs directory. Each experiment is a subfolder containing config.json, metrics.jsonl, and optionally eval.json. Can also be set with the --runs-dir command-line option. |
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
This server publishes no instructions, or was last inspected before Glama recorded them.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": false
} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| list_runsA | List all training runs with scene, planned iterations and status ('finished' has final metrics, 'no_eval' does not). Call this first if you don't know the run names. |
| get_run_configB | Get the training config of one run (iterations, learning rates, densification settings...). |
| get_final_metricsA | Get the final test metrics (psnr, ssim, lpips) of one run. Returns status 'no_eval' if the run has none. |
| compare_runsA | Rank several runs by one final metric and say which is best. Knows that lower lpips is better. |
| get_metric_curveA | Get a downsampled training curve of one logged metric (e.g. loss, train_psnr, num_gaussians) plus its last value. |
| diagnose_runA | Check one run for problems: missing final eval, NaN/inf loss, loss spikes, stopping early, too many Gaussians. Use it to explain why a run failed or has no results. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 6 tools
Each tool has a clearly distinct purpose: listing runs, reading config, reading final metrics, comparing runs, reading one metric curve, and diagnosing failures. The only apparent overlap between get_final_metrics and compare_runs is resolved by their descriptions: one reads a single run's metrics, the other ranks multiple runs.
All tools follow a consistent snake_case verb_noun pattern: list_runs, get_run_config, get_final_metrics, compare_runs, get_metric_curve, diagnose_run. There are no mixed conventions or vague names.
Six tools is well-scoped for inspecting and analyzing training runs. Each tool covers a distinct analysis need, and none feels redundant or extraneous.
The set covers the core lifecycle of run inspection: listing, configuration, final metrics, curve inspection, comparison, and diagnosis. Minor gaps remain, such as filtering runs or comparing metric curves across runs, but agents can work around these with the existing tools.