Feature · Agent + Drift
Bring your own model. Then drift-test what it touched.
MCPOrbit's Agent panel runs against any model provider you have credentials for. Use it to drive end-to-end tests, including drift detection: baseline a server's capabilities once, re-run later, and MCPOrbit highlights every schema change, new tool, or removed permission.
Providers
Eight providers, one panel.
Anthropic
Native Claude support, including the latest Opus and Sonnet.
OpenAI
GPT-5, GPT-5 Mini, the o-series reasoning models.
OpenRouter
One key, every router-supported model. Useful for cheap exploration.
AWS Bedrock
Claude, Llama, Mistral models hosted on Bedrock with your AWS credentials.
Groq
Sub-50ms first-token latency on Llama and Mixtral. Fast iteration.
Ollama
Fully local. Pull a model once; run it with no network.
LM Studio
Local model server with a GUI. Pair with MCPOrbit when you want both UIs.
Any OpenAI-compatible
Custom endpoint, custom model. Behind your VPC, in your test cluster, wherever.
Drift testing
Three moves: baseline, re-run, diff.
Capabilities of an MCP server change. Tool descriptions get rewritten. Parameters get renamed. Permissions get widened. The tools/listyou wrote your agent against six weeks ago might not be the one you're calling today. Drift testing makes the change visible before it ships.
- Baseline — save a snapshot of every tool, every schema, every permission scope
- Re-run — reconnect later, in a week, in a release cycle
- Diff — schema changes, new tools, removed tools, permission widenings, colour-coded by severity
drift · linear-mcp · vs baseline 2026-04-21
new tool: refund_order
args: order_id (string, required), reason (string)
schema changed: list_tickets
status enum widened: + 'archived'
removed tool: legacy_export
callers will receive MethodNotFound after 2026-05-15
Local-first
Plug Ollama or LM Studio in. Run the agent fully offline.
Useful when the MCP servers you're testing handle data you'd rather not put through a hosted model. Same agent UI, different runtime. Same drift workflow, against an air-gapped MCP server if you want.
- No network needed for inference
- Bring an air-gapped MCP server in too
- Same agent UI, different runtime
Agent · local
Ollama · llama-3.3-70b
> tools/call list_files { path: '/private/notes' }
Agent panel and drift FAQ
Where do I put my API keys?
Locally, in MCPOrbit's settings. Keys never leave your machine. There is no cloud account.
Do you store my prompts or responses?
No. Telemetry, when shipped, never includes payloads — only metadata. The full schema is on /privacy/telemetry once that pipeline goes live.
Can drift baselines be checked into git?
Yes. Each baseline is a JSON file MCPOrbit writes to a path of your choosing. Commit it, version it, share it.
Can I run drift in CI?
Not yet. The headless runner is on the roadmap. Today it's a developer tool, not a pipeline tool.
Multi-server agent runs?
Single-server agent runs ship today. Multi-server orchestration in one session is on the roadmap.
