Agent Breadcrumbs is a lightweight logging and observability tool for agent work, designed for cross-client and cross-platform use, covering Codex, Claude, Cursor, ChatGPT, OpenClaw, and other clients.
Purpose: It aggregates work completed by different agents through a single MCP tool (log_work), can target a single user or multiple users, and helps teams uniformly record and observe agents' work processes.
Main capabilities:
- Aggregate work records across different agents through the MCP tool log_work.
- Supports a custom schema, or use a built-in profile to define the log structure; once defined, the MCP schema guides clients to log with the correct payload, without needing to repeatedly explain the payload format on every tool call.
- Output sinks are selectable: local JSONL, Webhook, Postgres.
- Provides starter schema profiles for common use cases: agent insights, delivery tracking, audit trails, knowledge capture.
- Provides a simple dashboard app for viewing recorded work.
How to use:
- MCP quick start: You can install and run it directly with the default configuration, or run it with an explicit config file; --config is optional, and when omitted, server-side defaults are used.
- Client configuration examples include Codex (~/.codex/config.toml) and Claude Desktop (claude_desktop_config.json).
- After installing the MCP server, you need to use mcporter so that the server can be called via the CLI, and then add the system prompt in the channel you are using.
- For cron-driven workflows, you can instruct the agent to call log_work on each scheduled run to perform time-based regular logging.
- Repository quick start: You can run the MCP server locally with an example configuration, and you can also run the dashboard locally.
MCP configuration model: The top-level configuration file is JSON, and supports either schema (fully custom log_record properties) or schema_profile (use a built-in profile from packages/mcp/examples/schema_profiles); the two cannot be set at the same time. It also supports user_name (on the server side,inject user identity into persistent records) and sink (target settings: jsonl, webhook, postgres). Default behavior when omitted: schema is the built-in default (agent_id, timestamp, work_summary, additional); the output file is ~/.agent-breadcrumbs/logs.jsonl; user_name is not injected. Your schema/profile is applied directly to the MCP tool input schema, and agent clients can see the required fields of log_record from the tool definition.
Repository structure: MCP server package (published): packages/mcp; dashboard app (repository-only, unpublished): apps/dashboard.
Suitable scenarios: teams that need to unify logging standards across multiple agent clients and route data to the desired destinations; it can cover scenarios such as agent insights, delivery tracking, audit trails, and knowledge capture.
Limitations: The materials do not provide more information about pricing, cost, platform support scope, actual usage results, or performance, so these aspects are not provided.
