Gateplex sits inline between your agents and the systems they touch. Every action is evaluated against your rules first, resolved to a verdict, and written to a tamper-evident record.
| Time | Agent | Attempted action | Rule evaluated | Verdict |
|---|---|---|---|---|
| 14:02:44.812 | payments-reconciliation | Initiate wire transfer, 148,200.00 GBP | Spend Limit, 100k GBP | BLOCK |
| 14:02:41.304 | mortgage-drafting | Generate customer offer letter | Keyword Flag, disclosure set | ALLOW |
| 14:02:39.118 | kyc-review | Read customer record, national ID field | PII Shield | PENDING_APPROVAL |
| 14:02:35.550 | collections-outreach | Send message to customer segment 4 | Output Length Limit | FLAG |
| 14:02:31.207 | treasury-sweep | External call, unlisted vendor domain | API Scope Control | BLOCK |
There is no language model in the decision path. The same action, evaluated against the same rules, returns the same verdict every time.
| Verdict | Effect on the action | Record |
|---|---|---|
| ALLOW | Proceeds. | Logged with the evaluation result. |
| BLOCK | Stopped at the boundary. | Logged with the rule that stopped it. |
| FLAG | Proceeds, and is marked for review. | Surfaced in the live feed. |
| PENDING_APPROVAL | Held. The agent waits for a decision. | Queued to a named reviewer, decision recorded against their identity. |
| Rule | What it governs |
|---|---|
| Spend Limit | Caps the monetary value an agent can move in a single action or window. |
| PII Shield | Detects personal data in an agent's input or output before it moves. |
| Keyword Block | Stops actions containing prohibited terms. |
| Keyword Flag | Records actions containing sensitive terms without stopping them. |
| Rate Limiter | Constrains how frequently an agent can act. |
| Rule | What it governs |
|---|---|
| API Scope Control | Restricts which external endpoints and domains an agent may reach. |
| Regex Pattern Match | Matches structured identifiers such as account or card formats. |
| Output Length Limit | Bounds the size of generated output. |
| Prompt Injection Detector | Identifies attempts to override the agent's instructions. |
| Repetition Detector | Catches looping or repeated agent behaviour. |
Preconfigured rule sets for the regimes our customers are held to, editable once applied.
Rules configured to require approval return PENDING_APPROVAL. The action is held until a named reviewer approves or rejects it, and their identity is derived server side, not supplied by the client.
Read customer record, national ID field
PII Shield · raised 14:02:39 · execution held
Approve facility above delegated authority
Spend Limit, 250k GBP · raised 13:58:04 · execution held
A single intercept endpoint. Send the action, receive the verdict, act on it.
The gateplex-python package on PyPI, including drop-in patches for the OpenAI and Anthropic clients.
Published on Smithery, Glama, and MCP.so for agents that speak the Model Context Protocol.