Guardrails for AI coding agents — blocks catastrophic tool calls before they run

Deterministic rule engine: materializes per-session context and guards tool calls.

Deterministic rule engine: materializes per-session context and guards tool calls.

Cordon for MCP — security gateway for MCP tool calls. Firewall, auditor, and human-in-the-loop approvals over any stdio MCP server.

Convert chat conversations and responses between OpenAI, Anthropic and Gemini. Tool calls, images, audio, documents and roles handled. Zero dependencies.

Deterministic rule engine: materializes per-session context and guards tool calls.

Deterministic rule engine: materializes per-session context and guards tool calls.

CLI for reading local Codex session rollouts and extracting assistant replies, tool calls, and MCP activity

Trace plugin for Electron Stagewright: record a driving session to JSONL, report token usage, and replay the recorded tool calls.

Detect and break agent doom loops in pi: blocks identical repeated tool calls and blind retries before they burn tokens.

Tool-input repair layer for the pi coding agent: validate-then-repair for built-in tool calls, ported from the behavior of commandcode's repair layer

Agent runtime for the Graphorin framework: createAgent({...}) factory, the typed model -> tool calls -> model loop, the AgentEvent<TOutput> discriminated event stream, steering / followUp queues, durable HITL approvals via runStateToJSON / runStateFromJSO

Pi coding agent extension that runs all tool calls inside a remote Upstash Box sandbox while the agent runs locally

Agent behavior tracking SDK for Cairo — generations, tool calls, decisions, errors

Deterministic rule engine: materializes per-session context and guards tool calls.

Runtime security SDK for AI agents — guard tool calls in 1 line

Adapter-driven chat surface — drop-in <Chat>, composed render-prop, primitives, segment-aware tool calls, ask-user, sessions, slash trigger, settings popover, voice button.

Find tool calls that could fire the same effect twice. Reports evidence, renders no verdict. Zero dependencies.

Catch leaked secrets, prompt injection, and destructive calls in your AI agent's tool calls. Runs locally with no signup; enforce centrally with AxioRank.

Record Model Context Protocol tool calls as tamper-evident traceglass runs