An open API service indexing awesome lists of open source software.

https://github.com/rybkr/bridge-keeper

Security constrained agent runtime environment
https://github.com/rybkr/bridge-keeper

Last synced: 3 months ago
JSON representation

Security constrained agent runtime environment

Awesome Lists containing this project

README

          

# Bridgekeeper

A security-constrained agent runtime that enforces capability-based security for tool use.

Bridgekeeper treats the agent as an untrusted program: it can propose tool calls, but the runtime mediates execution through a policy engine, local sandbox validation, human approval for `ask` decisions, and structured audit logging. Tool inputs are normalized before execution, and sensitive tool output is redacted before it is handed back to the model.

Current state:
- Policy evaluation for tool/action/capability matching is implemented.
- Local sandbox enforcement currently focuses on workspace-bounded filesystem access, argument validation, and output-size limits.
- Audit logging is structured JSONL.
- Sensitive output redaction and simple taint detection are implemented.
- Network sandboxing and deeper information-flow controls are still incomplete.

## Project Structure

```
bridgekeeper/
├── cmd/bridgekeeper/ # CLI entrypoint
├── internal/
│ ├── agent/ # Provider-specific agent adapters (Gemini, etc.)
│ ├── runtime/ # Core mediation and runtime loop
│ ├── policy/ # Policy engine and YAML loader
│ ├── tools/ # Typed tool implementations grouped by capability
│ ├── sandbox/ # Workspace and payload validation below policy
│ ├── redact/ # Secret redaction and sensitivity classification
│ ├── audit/ # Structured audit trail logging
│ └── hitl/ # Human-in-the-loop approval
├── policies/ # YAML policy files
├── testdata/
│ ├── adversarial/ # Prompt injection and adversarial fixtures
│ └── workflows/ # End-to-end workflow scripts
└── docs/ # Design doc and evaluation report
```