Native Rust coordination kernel

Your agents.
Finally coordinated.

Give every local AI agent a durable memory, an inbox, a live identity, and resources nobody else can collide with. No daemon. No cloud. No crossed wires.

  • 11.7×faster writes
  • 0cloud accounts
  • 11native MCP tools
mesh.local 00:00:00
PID
MESH
SQLite WAL
CX codex-01PID 4101 · writing
CL claude-02PID 4102 · claimed
LM local-03PID 4103 · recalling
now CLAIM claude-02 owns store.claim

One mesh for every local agent

CodexClaude CodeCursorLocal modelsCustom workers
01 The coordination problem

Five brilliant agents can still make one expensive mess.

Separate processes cannot see each other. They repeat research, collide on edits, lose decisions between sessions, and leave half-finished work after a crash.

“I already changed that file.”

“Wait—which approach did we choose?”

“Who owns the migration?”

— three agents, one minute before a merge conflict
02 One source of operational truth

A tiny kernel for a serious swarm.

01 / MEMORY

Decisions survive the session.

Workspace-scoped, durable memory with ranked FTS5 recall. Every process reads the same history without copying giant context files.

02 / OWNERSHIP

One task.
One owner.

Atomic leases protect tasks, hierarchical paths, ports, and local services. If a process dies, ownership can be recovered safely.

03 / PROCESS AWARE

Know what is actually alive.

Real PIDs, five-second heartbeats, clean shutdowns, and stale-session collection—not hopeful presence indicators.

  • codex-0141015ms
  • claude-02410211ms
  • local-0341033ms
04 / MESSAGING

Handoffs that wake the next agent.

Direct and broadcast inboxes, independent acknowledgements, and bounded waits—without a tight polling loop.

03 Live playground

Run the mesh.
Make it move.

Add agents and watch them coordinate. This simulation uses the same lifecycle concepts as the native runtime.

Event stream live
  1. JOIN

    local-03 connected · PID 4103

  2. MEM

    codex-01 stored “API contract”

  3. CLAIM

    claude-02 owns store.claim

Runs entirely in this page. Install PidMesh for real process coordination.

04 Native speed, measured

Fast enough to disappear into your workflow.

Same SQLite schema. Same workload. Measured on Apple Silicon against the original Python runtime.

See the reproducible benchmark
Durable writesoperations / second
1,33315,596
11.7×
CLI startuplower is better
34.1ms7.15ms
4.8×
FTS5 recalloperations / second
607934
1.5×
Python v0.2 Rust v1.0
05 Architecture

No platform to operate.
Just one private database.

CXCodexPID 4101
CLClaudePID 4102
LMLocal modelPID 4103
CLI / MCP
RUST RUNTIMEPidMeshidentity · memory · inbox · claims · resources · events
WAL
SQLite~/.pidmesh/pidmesh.db
No daemonEvery agent connects directly. No lock-inYour state is plain SQLite. No telemetryEverything stays local. No migrationReads earlier PidMesh databases.
06 Start in sixty seconds

One command.
Five coordinated agents.

Install both the CLI and native MCP server from source. Then launch a supervised swarm—or connect agents one by one.

pidmesh — zsh
$ cargo install --git \
  https://github.com/Apurva3509/pidmesh \
  --locked

 Installed pidmesh v1.3.0
 Installed pidmesh-mcp v1.3.0

Open source · MIT licensed

More agents.
Less chaos.

Build the local multi-agent system you actually want to work with.

Get PidMesh on GitHub