Intercepts tool calls at the boundary
Phinq observes agent tool calls before execution rather than relying on prompts or instructions the agent could ignore.
Phinq
ReclamarPhinq is an open source runtime governance layer for AI agents. It intercepts agent tool calls, classifies them by risk, and pauses higher-risk actions for human approval before execution.
Runtime governance for AI agents
Phinq is an open source runtime governance layer for AI agents. It sits between an agent’s proposed tool call and the execution boundary, classifies the request by risk, and either allows it to pass or holds it for approval.
The product is meant to reduce damage from irreversible agent actions such as deletes, credential reads, payments, or external messages. It records each decision in an append-only, hash-chained audit log so teams can review what the agent attempted and how the request was resolved.
Core capabilities
Classifies requests by risk
The system evaluates the proposed action using the tool name, arguments, and session velocity, then returns an allow or hold decision.
Holds irreversible actions for approval
Higher-risk actions are paused instead of forwarded until a human approves them through the configured channel.
Fails closed on unanswered holds
If a held action is denied or times out, the action is blocked and never reaches the tool.
Records a tamper-evident audit chain
Every decision is written to an append-only JSONL log with SHA-256 chaining so later edits, reordering, or deletion can be detected.
Runs in multiple integration modes
Phinq can run as a proxy, an in-process SDK, or a plain HTTP gate, and the docs list compatibility with OpenAI-style, Anthropic Messages, Gemini, and MCP-based clients.
Practical use cases
Approve dangerous agent actions before they execute
Use Phinq when an assistant may need to delete records, read credentials, send payments, or dispatch messages, but you want a human decision point before the action reaches the tool.
Add governance to existing agent stacks
Teams can place Phinq at the proxy, SDK, or HTTP boundary instead of rewriting the agent workflow, which makes it suitable for systems already using compatible clients or MCP.
Keep an audit trail of agent decisions
Phinq is useful when you need to review what an agent attempted, how it was classified, and whether it passed, held, approved, denied, or timed out.
Start in observe-only mode before enforcement
The docs say the proxy starts in shadow mode by default, which lets teams classify and log traffic before enforcing holds in production.
Pros and Cons
Pros
- Intercepts actions before execution, which is more enforceable than prompt-based guidance.
- Separates safe actions from risky ones so routine work can continue without manual review.
- Provides a tamper-evident audit log for post-incident review and verification.
- Supports multiple deployment styles, including proxy, SDK, and HTTP gate.
- Works with several documented client ecosystems, including OpenAI-style, Anthropic Messages, Gemini, Claude Code, Codex, Gemini CLI, LangChain, CrewAI, Mastra, and MCP.
Cons
- Risky actions require a human decision, so some workflows will pause.
- The documentation says the proxy starts in shadow mode by default, so enforcement must be enabled deliberately.
- The audit log does not include tool-call arguments or message payloads; those live in a separate corpus that is not hash-chained.
FAQ
What does Phinq do?
Phinq is a runtime governance layer for AI agents that intercepts tool calls, classifies them by risk, and holds higher-risk actions for human approval before execution.
How does Phinq decide whether to allow an action?
The docs say it evaluates the proposed tool call using the tool name, arguments, and session velocity, then returns allow or hold.
Does Phinq enforce decisions by default?
No. The documentation says the proxy starts in shadow mode by default, meaning it classifies and logs traffic first and only enforces HOLD decisions when enforcement is enabled.
What happens when a risky action is held?
The action becomes a pending approval instead of being forwarded. It can then be resolved through the configured channel; denied or timed-out holds are blocked.
Is the audit log tamper-evident?
Yes. The audit log is append-only JSONL and each entry is chained with SHA-256, so changing, reordering, or deleting entries breaks verification.
Quick Facts
- Category
- Developer Tool
- Product type
- Open source runtime governance layer
- Primary function
- Classifies AI agent tool calls and pauses risky actions for approval
- Deployment modes
- Proxy, TypeScript SDK, plain HTTP gate
- Documentation domain
- phinq.co
- License
- MIT