# Pulsar gives local coding agents a shared discovery and messaging bus
An open-source project called Pulsar has introduced a local messaging layer intended to let AI coding agents discover and communicate with one another even when they run in different harnesses. The MCP server currently lists support for Claude Code, Codex, OpenCode, pi, omp, Cursor and Gemini CLI.
Pulsar places all participating agents on one machine-wide bus. Agents can announce what they are working on with tags, inspect which other agents are available and send a message to one recipient, a tagged group or every participant. The project says an incoming message can also wake an agent that is idle, depending on the receiving harness’s supported delivery mechanism.
The design uses one lightweight daemon per machine to maintain the registry and a SQLite message log. Each agent connects through a thin MCP client over a Unix socket restricted to the local user. The daemon starts when needed and stops after becoming idle, avoiding a permanently running service.
Installation is designed to accommodate the different configuration formats used by agent tools. Pulsar’s setup command searches for supported clients, shows which files it intends to modify and asks for confirmation. It creates backups of changed files and installs a versioned local copy under the user’s data directory. A stable link points harnesses to the current version so a later setup run can update the installation without making every agent session wait for a package download.
Users may instead install through a harness’s own package mechanism, although the project notes that some capabilities require the full setup. In particular, push delivery extensions for pi and omp use the npm package, while Claude cross-session delivery and Codex tool-approval settings need configuration beyond a basic plugin install. When a client cannot accept pushed messages, Pulsar retains them in an inbox and reports unread items in later tool results.
The daemon also attempts to identify live Claude Code and OpenCode sessions that began without Pulsar loaded, allowing them to appear in discovery and receive messages through their harness adapters. That behavior is meant to reduce fragmentation when only some running agents were launched with the MCP server already configured.
Pulsar’s present scope is communication among agents on a single computer. The maintainers describe multi-machine messaging and remote control over a Tailscale connection as future work, rather than shipping features. The current version should therefore be viewed as local coordination infrastructure, not a general remote-agent network.
The release reflects a broader shift toward multi-agent development workflows, where separate tools may research, implement and test parts of the same task. Pulsar does not decide how work is divided or validate another agent’s output. Its role is narrower: provide identity, presence and message transport across otherwise isolated clients.



