Gravitre AI (/ai) is the unified AI surface. Execute mode (/ai?mode=execute) is for assigned, monitored work: describe an outcome; Gravitre builds an action plan, surfaces context from Runs and Connectors, and tracks async jobs with explicit verification labels. It is not a free-form chat thread.
Legacy routes /command-center and /operator redirect to Gravitre AI.
Open Execute mode
Sidebar → Gravitre AI (/ai), then select Execute mode — or open /ai?mode=execute directly.
The page shows:
- Command input — your instruction and optional context attachment
- Environment badge — staging vs production scope
- Execution mode — how aggressively Gravitre may invoke tools
- Model selector — Auto or explicit LLM routing (including GIBE-backed routing when configured)
- Timeline — findings, plan steps, and job status
- Suggested actions — one-click follow-ups with confidence and approval badges
Gravitre AI modes compared
| Execute | Chat | Find / Search | |
|---|---|---|---|
| Goal | Run a plan with tracked jobs | Answer questions in dialogue | Locate records with deep links |
| Output | Action plan + Run linkage | Conversational reply | Grouped search results |
| Route | /ai?mode=execute | /ai?mode=chat | /ai?mode=find or /search |
Use Execute when you want durable work with a visible plan. Use Chat for exploration and Q&A. Use Find when you know what you are looking for and need a link — not a chat reply.
Typical flow
Attach context (optional) → Enter goal → Review action plan → Approve gated steps → Monitor job → Read verification summary
Step 1 — Attach context (recommended)
Pin context so the plan is grounded in real org data:
| Context type | When to use |
|---|---|
| Run | Debug or extend a failed/successful run |
| Workflow | Change or troubleshoot a specific automation |
| Connector | Fix sync errors or credential issues |
| Source | Query or ingest related knowledge |
Context packs load via /api/operator/context/{type}/{id} and appear as chips in the Gravitre AI UI.
Step 2 — Write a strong instruction
Include goal, systems, constraints, and approval preferences:
In production: when a PagerDuty incident resolves, post a summary to #incidents in Slack.
Do not post until I approve the draft message.
Review failed workflow runs from the last 24 hours for the HubSpot sync workflow.
Propose fixes but do not change workflows automatically.
Step 3 — Review the action plan
The plan lists numbered steps with status (pending, current, completed). Check:
- Required Connectors — connect missing integrations before approving execution
- Approval badges — steps that write to external systems may require human sign-off
- Confidence labels — GIBE-backed suggestions show confidence when available
Step 4 — Approve and monitor
Approve gated steps in-line or from Approvals (/approvals). Monitor job progress in the timeline; open linked Runs for step-level detail.
Step 5 — Read the verification summary
The summary distinguishes what was analyzed, proposed, and executed. Skipped steps usually mean a missing Connector, failed health check, or pending approval.
Connector availability
Before execution, Gravitre checks connector readiness: Configured → Authenticated → Healthy → Executable. The same checks appear on the Connectors page and in Chat mode tool chips.
What GIBE contributes
Execute-mode plans can incorporate GIBE signals — failure predictions, retrieval quality, and org memories — with confidence scores shown in the UI. Recommendations stay advisory until you approve.
API
Context and action-plan endpoints (stable paths):
curl -s "https://gravitre.app/api/operator/context/run/RUN_ID" \
-H "Authorization: Bearer YOUR_API_KEY"
curl -s "https://gravitre.app/api/operator/context/workflow/WORKFLOW_ID" \
-H "Authorization: Bearer YOUR_API_KEY"
curl -s -X POST "https://gravitre.app/api/operator/action-plan" \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"instruction": "Summarize failed runs today", "environmentId": "ENV_ID"}'
See OpenAPI for schemas.