MCP Install
Connect the AXP MCP server to Cursor, Claude, ChatGPT, Codex, or another MCP host.
The AXP MCP server lets agents query and share results, manage experiments and
runs, create guided trace analyses and findings, manage org-scoped secrets,
switch org context, and authorize the local ax CLI with ax auth connect.
Use the install instructions for your agent host:
Install the AXP MCP server directly in Cursor. If Cursor opens MCP settings without an install prompt, add this same server config manually.
{
"mcpServers": {
"AX": {
"url": "https://app.514.ax/mcp"
}
}
}After connecting the MCP server, follow any sign-in prompt from Cursor, Claude, ChatGPT, Codex, or your MCP host. Then verify AXP tools are available in your agent session.
Except for org-context tools and the CLI auth bootstrap, org-scoped tools use the
current MCP org unless you pass an org_id override.
Tool catalog
| Group | Tool | Purpose | CLI equivalent |
|---|---|---|---|
experiment | experiment_list | List experiments the current org can access. | ax experiment list |
experiment | experiment_view | Show one experiment's metadata, latest version, axes, and run counts. | ax experiment view |
experiment | experiment_versions | List stored YAML version fingerprints for an experiment. | ax experiment versions |
experiment | experiment_pull | Return the stored YAML for the latest or selected experiment version. | ax experiment pull |
experiment | experiment_push | Validate and push a new experiment YAML version without starting runs. | ax experiment push |
experiment | experiment_run | Submit runs for the latest pushed version of an experiment. | ax experiment run <experiment-id> |
run | run_list | List individual runs with optional experiment, version, dimension, status, and limit filters. | ax run list |
run | run_view | View a run request roll-up or one composite run id. | ax run view |
run | run_cancel | Cancel queued or running work for a run request, optionally limited to specific run UUIDs. | ax run cancel |
results | results_view | View modeled experiment metrics with experiment_id, or one run's metrics with run_id. | ax results view --experiment <id> / ax run view <run-id> |
results | results_query | Run read-only SQL against uploaded AXP Source/Facts tables for novel questions. | ax results query |
results | results_share | Mint or reuse a public results-explorer share URL for an experiment. | ax results share |
results | results_dimensions | List valid results filter and group-by dimensions and values for an experiment. | ax results dimensions |
results | results_scatter | Return a deterministic sample of run cost/time scatter points for an experiment. | ax results scatter --experiment <id> |
secret | secret_set | Create or overwrite one org secret; returns metadata only, never the value. | ax secret set |
secret | secret_list | List secret slugs and audit metadata; never returns values. | ax secret list |
secret | secret_delete | Delete one secret by slug. | ax secret delete |
analyze | analyze_suggest | Suggest guided trace-analysis questions for failed or error runs. | ax analyze suggest |
analyze | analyze_prompt | Build a copy/paste prompt for the full trace-analysis workflow. | ax analyze prompt |
analyze | analyze_create | Validate and persist caller-supplied trace analysis output. | ax analyze create |
analyze | analyze_get | Read one persisted trace analysis. | ax analyze get |
analyze | analyze_list | List persisted trace analyses. | ax analyze list |
findings | findings_resolve | Resolve run, tool call, timeline entry, span, or prompt anchors into trace evidence. | ax findings resolve |
findings | findings_create_analysis | Create the analysis record used by a saved findings result. | ax findings create --question --result |
findings | findings_create | Create a persisted finding with quoted evidence links. | ax findings create |
findings | findings_list | List persisted findings for a saved result. | ax findings list |
findings | findings_get | Read one persisted finding with evidence links. | ax findings get |
mcp_org | mcp_org_list | List organizations available to the authenticated MCP user and show the selected org. | ax auth orgs --list |
mcp_org | mcp_org_get_current | Get the current organization context used by org-scoped MCP tools. | ax auth orgs |
mcp_org | mcp_org_set_current | Set the organization context used by org-scoped MCP tools. | ax auth orgs --switch <org-id> |
cli_auth | cli_auth_start | Create a short-lived one-time command that authenticates the local ax CLI. | ax auth connect --code <code> |
Start trace work with analyze_suggest for a failed/error run batch; it returns
the structured prompt, ax results query calls, and create-request skeleton to
use before persisting with analyze_create.
Results scatter tool
results_scatter is read-only. Pass required experiment_id; pass optional
org_id only when you want to override the current MCP org for that call. The
tool returns both text and structured data channels: text for a compact human
summary, and structured JSON for clients that render charts. MCP hosts that
support interactive MCP-Apps widgets can show the same points as an embedded
scatter plot widget; hosts without widget support still receive the text and
structured data.
The point sample is deterministic and capped at 2,000 runs. The response includes
totalPoints and truncated so callers can disclose when the widget or JSON is
showing a sample instead of every matching run.