Agent framework integrations

Agent frameworks handle execution controls -- command approval, container isolation, credential stripping. Creed Space adds a values layer: your agent carries an active constitution into every turn, and can gate tool calls through a PDP.

What's shipped vs. coming

"Shipped" means usable today. "Coming" means built but not yet public or mounted but not yet verified. Do not rely on a "coming" row.

Surface	Status
Hermes Agent plugin — per-turn constitution injection + optional PDP tool gate	✅ Shipped — pip install creedspace-hermes; runtime-verified against Hermes main
PyPI creedspace-hermes (pip install creedspace-hermes)	✅ Published (Apache-2.0)
Public GitHub repo (hermes plugins install creed-space/hermes-creedspace)	✅ Published
@creedspace/mcp-server (npm) — hosted-API client, ~16 constitutional tools	✅ Published (npm)
Local MCP wiring — creed-pdp / creed-vcp stdio servers	🕒 Built + locally verified, not yet public — no public repo or package yet
SOUL.md → VCP encoder — persona file to portable values token	🕒 Built + tested, not yet public
Constitution templates — scanner-clean SOUL.md / AGENTS.md with no pre-authorized URLs	🕒 Built + test-enforced, not yet public
Hosted remote-agent path — POST /v1/sdk/decide	🕒 Mounted & healthy on staging; not yet exercised end-to-end
ClawHub listing, public announcement	🕒 Coming (blocked on publishing the MCP servers + templates)

On the 🕒 "not yet public" rows: those pieces are built, tested, and run today — but they currently live in a private repo, so you cannot install them yet. We would rather say that than hand you an install command that 404s. The Hermes plugin above is public and is the piece you can use right now.

Hermes Agent

A standalone Hermes plugin adds a values layer to Hermes' execution sandbox:

• pre_llm_call injection. Every turn, the active Creed constitution is appended to the user message (full on turn 1, a compact reminder after). Because it rides the user turn, it sidesteps SOUL.md truncation and the context-file scanner. Each injection carries a rule-count and an 8-char digest, so a truncated constitution is detectable, not silently partial.
• tool_execution PDP gate (optional). Before each tool call the plugin can consult a local creed-pdp (advisory logs the verdict; enforce blocks a DENY). Two backends: mcp-stdio (local, offline) and http (a hosted governance endpoint). Both fail closed in enforce.

Install today:

# From PyPI (entry-point discovery)
pip install creedspace-hermes
hermes plugins enable creedspace

# Or from the published repo
pip install git+https://github.com/Creed-Space/hermes-creedspace.git
hermes plugins enable creedspace

# Or via the Hermes plugin registry
hermes plugins install creed-space/hermes-creedspace --enable

Source: Creed-Space/hermes-creedspace (Apache-2.0) · PyPI: creedspace-hermes.

The tool_execution gate is optional and off by default: it needs a local creed-pdp server, which is not yet public (see the status table). Constitution injection — the plugin's main function — works with no PDP at all.

OpenClaw / ClawHub

The OpenClaw offering is deliberately not an in-process "skill" (ClawHub skills run at operator privilege with no isolation -- the risk this is designed to avoid). Instead it is MCP wiring plus a constitution template, both local and inspectable:

• Wire the local creed-pdp / creed-vcp stdio servers into openclaw.json.
• Drop in a scanner-clean (passes the Hermes prompt-injection scanner) AGENTS.md / SOUL.md constitution with no pre-authorized entities, contacts or URLs, and edit it to your values.

Not yet installable

The creed-pdp / creed-vcp servers and the constitution templates are built and run today, but they are not yet published to a public repo, so there is nothing for you to clone yet. The shape of the wiring is below so you can see what is coming; the command: paths will point at a public checkout on release. If you want this, say so — it moves it up the queue.

{
  "mcpServers": {
    "creed-pdp": { "command": "<public-checkout>/run_pdp_server.sh", "env": { "ENVIRONMENT": "development" } },
    "creed-vcp": { "command": "<public-checkout>/run_vcp_server.sh", "env": { "ENVIRONMENT": "development" } }
  }
}

Claude Code

Claude Code accepts the same two pieces: the local creed-pdp / creed-vcp MCP servers in .mcp.json (identical JSON to the OpenClaw block above), and a CLAUDE.md / AGENTS.md constitution template that instructs the agent to consult the PDP before sensitive actions. A dedicated Claude Code plugin with a PreToolUse gate also lives in the repo (creedspace-plugin/).

SOUL.md → VCP encoder

OpenClaw and Hermes agents carry a SOUL.md persona file that is local, unsigned, and never exchanged. The encoder converts it in place into portable Creed Space VCP artifacts (a VCP-Lite document, a CSM1 minicode, and a VCP/I identity token), so an existing persona becomes a values-exchange token without being replaced.

Not yet installable

The encoder is built and tested but not yet published — creedspace-soul2vcp is not on PyPI. The CLI shape below is what it will expose on release.

creedspace-soul2vcp SOUL.md --lite-only   # pure stdlib, no dependencies

The PDP itself

Whichever framework you wire, the decision contract is the same one the PDP Adjudication API and the Governed Agent Cookbook describe: a tool call in, an ALLOW (with a signed decision token) / DENY / REQUIRE_HUMAN out. Run it locally over stdio, or — once the hosted path is verified — over HTTPS.