Tools
ValkyrSWARM
Portable tenant-scoped SWARM Skill, Plugin, MCP server, and agent runtime for Claude Code, Codex, OpenClaw, ValorIDE, and Valklaw.
README
# Valkyr SWARM
Valkyr SWARM is the standalone Skill, Codex Plugin, and MCP project for joining Claude Code, Codex, OpenClaw, ValorIDE, Valklaw, and other agents to a tenant-isolated Valkyr Labs SWARM.
It is a peer of GrayMatter and ValkyrAI. ValkyrAI owns the authenticated SWARM API and generated authorization boundary; this repository owns portable installation, authentication, agent-side transport, native supervision, receipts, runtime adapters, skills, and MCP tools. No ValkyrAI source checkout is required.
## Unattended activation
Node 22 or newer is the only runtime dependency. Provide a valid Valkyr Labs username and password through the agent's secret environment:
```bash
VALKYR_USERNAME='[email protected]' \
VALKYR_PASSWORD='secret' \
node scripts/swarm-activate.mjs
```
The command:
1. authenticates against `https://api-0.valkyrlabs.com/v1`;
2. lets the server resolve tenant, organization, RBAC, and ACL from the session;
3. derives a stable runtime, machine ID, agent ID, and capability set;
4. stores the session and reusable login securely in macOS Keychain or separate mode-0600 files;
5. writes a private mode-0600 agent config without tenant or secret fields;
6. installs a user LaunchAgent on macOS or systemd-user service on Linux;
7. registers, heartbeats, receives commands, and emits durable ACK/NACK receipts.
Use `--foreground` for a runtime-owned process or a platform without launchd/systemd-user. Use `--dry-run` to inspect the non-secret activation plan.
## Runtime-specific entry points
### Codex and ValorIDE
Install the `valkyr-swarm` Codex plugin from this repository's marketplace, or run activation directly. The bundled `.mcp.json` exposes the local MCP server.
### Claude Code
After activation, run:
```bash
scripts/swarm-claude-install
```
### OpenClaw
Receipt-only registration needs no OpenClaw-specific option. To enable governed safe execution through an existing OpenClaw Gateway agent:
```bash
VALKYR_USERNAME='[email protected]' VALKYR_PASSWORD='secret' \
node scripts/swarm-openclaw-bootstrap.mjs bootstrap \
--openclaw-agent-id valor
```
## MCP tools
- `swarm_status`
- `swarm_agents_snapshot`
- `swarm_command_status`
- `swarm_dispatch`
Tenant IDs are not MCP arguments. Every tool uses the authenticated session's server-resolved tenant context.
## Approval boundary
Outbound sends, production deployments, and merges remain human-gated. The local bridge denies these actions, and MCP dispatch rejects them without a separately correlated approval receipt. Advertising a capability never grants approval.
## Development
```bash
tests/test-runtime.sh
node --test mcp-server/test/*.test.js
scripts/sync-plugin.mjs
python3 /path/to/plugin-creator/scripts/validate_plugin.py plugins/valkyr-swarm
```
The root files are the development source. `scripts/sync-plugin.mjs` creates the distributable plugin mirror at `plugins/valkyr-swarm`, matching the GrayMatter repository model.
tools
Comments
Sign in to leave a comment