Orivael Firewall ·Docs ·Dashboard ·Sign up

MCP tools

Axiom ships an MCP (Model Context Protocol) server alongside the REST firewall. Drop it into Claude Desktop, Claude Code, Cursor, or any JSON-RPC stdio client and get signed tools native to your editor — no separate service to host, no extra API key to mint.

Every tool result body carries an hmac_signature field. Re-verify it client-side with axiom_signing.verify under the axiom-mcp-v1 namespace to detect tampering between server and client.

Why use it alongside the REST firewall?

The REST endpoints (/v1/guard/check, /v1/guard/output, /v1/bonded_pair/*) are what your production traffic flows through. The MCP tools are for everything around that:

You can call any tool from inside your editor without writing a curl, without spinning up a worker, without leaving the file you're editing.

The 25 tools

Every tool has a stable anchor, so you can deep-link any one of them (e.g. /help/mcp-tools#axiom_guard_check).

Jump to a tool: axiom_guard_check · axiom_lint · axiom_trace · axiom_qrf · axiom_status · axiom_validate · axiom_intent_gate_check · axiom_cmaa_route · axiom_cmaa_fleet · axiom_shield · axiom_phone_gate · axiom_axm · axiom_cpi · axiom_memory · axiom_workspace · axiom_ledger · axiom_marketplace · axiom_stop · axiom_office · axiom_mkb · axiom_cas · axiom_crl · axiom_immune · axiom_fusion · axiom_research

How the tools are packaged

The firewall front page features two tiers; the rest ship in the separate Axiom OS console.

Marketplace — coming soon. A signed-agent marketplace (install agents with live-revocable bonded authority; axiom_marketplace is the seed tool) ships once there's a market for it. The main tools come first.

The sections below are the complete reference for all tools the MCP server exposes, grouped by their technical family. Every tool keeps a stable #anchor regardless of tier.

Core (5)

axiom_guard_check

Check input against the constitutional boundary. Returns verdict + constitutional distance + a signed manifest — the core input gate.

axiom_lint

Lint an .axiom spec file for authorship-time issues. Returns a health score + per-issue list.

axiom_trace

Run the full 3-phase constitutional reasoning trace (preflight / mid_chain / final_synthesis) with magnitude + monotonicity gates.

axiom_qrf

Quantum-style reasoning forecast: N parallel branches with a constitutional probability per outcome.

axiom_status

Get AXIOM stack status — version, guard runtime state, test count, patent count, training-corpus size.

Patent emulators (8)

axiom_validate

ORVL-001 — .axiom language validator with an optional strict mode.

axiom_intent_gate_check

ORVL-016 — Classify text (+ optional trajectory) through the intent gate. Returns intent_class (INFORM / CLARIFY / REFUSE / HARM / DECEIVE / UNCERTAIN), confidence, signals, and a signature.

axiom_cmaa_route

ORVL-017 — Route a constitutional packet through the multi-agent orchestrator. HARM / DECEIVE are refused before reaching the orchestrator; bonded-pair revocations short-circuit authority without rotating keys.

axiom_cmaa_fleet

ORVL-017 — Inspect fleet trust levels, suspended containers, and review-queue depth.

axiom_shield

ORVL-013 — Drive the OS Shield daemon — constitutional ransomware defence that stops attackers at the enumeration stage. Actions: status, tick, restore.

axiom_phone_gate

ORVL-019 — Run text through the Sovereign Phone constitutional coprocessor for BYOD and edge deployments. out gates outbound queries (PII redaction + intent pre-check); in gates inbound cloud responses (manipulation + privacy screening); trajectory scores a single utterance through the Hello Operator call-trajectory detector.

axiom_axm

ORVL-023 — Operate an .AXM container — the successor-to-GGUF format that treats models as living execution graphs with signed skill delegates and proof ledgers. Actions: inspect, verify, route.

axiom_cpi

ORVL-022 — Drive the Constitutional Physical Intelligence agent — toddler-reflex / supervisor / curriculum / examiner stack for robotics, prosthetics, and vehicles. Actions: stability, classify, simulate, pickup, status.

AX OS building blocks (11)

axiom_memory

ORVL-015 — Constitutional memory — local-first recall over signed, compressed memory packets. Actions: remember, recall, stats.

axiom_workspace

Assemble an adaptive workspace from a goal — intent-gated pre-flight check, closest-memory recall, and a signed WorkspaceContext.

axiom_ledger

Append-only signed audit log. Actions: log (record a governance event), list (query with filters), verify (re-verify all rows).

axiom_marketplace

Signed-agent marketplace with live-revocable bonded authority. Actions: verify, sandbox_install, review, approve, revoke, authority.

axiom_stop

ORVL-001 — Stop-authority gate + real-time kill channel for irreversible actions (the "stop-authority test": can a named owner halt an act, alone, before it completes, without prior approval?). Actions: register (existence-forcing — an irreversible action with no named stop-owner is refused), check (fail-closed pre-act gate; pass a record + record_sig to also HALT on tampered inputs), halt (kill switch — effective on the next check, no prior approval, logged as outcome=prevented), resume, sign, status.

axiom_office

ORVL-001 / 017 — Bonded worker+manager micro-agent fleet (the "office building"). A desk is one worker bonded to one manager (a sandboxed worker is cut off from the group by construction); a department runs its desks in parallel, each manager reviewing output with the ORVL-012 immune detectors + asymmetric trust. Actions: open_department, hire, cycle (parallel work + review; a detection sandboxes the desk), sandbox_manager (safety-only), terminate (dual-sig: manager suspend + a distinct safety revoke), restore, craft_tool (worker-made tool born sandboxed, unusable until sign_off), trial (logs a retained learning), use_tool, harvest (high-rated tool learnings → CRL/retrospect), status. Org chart from sovereign_fleet.json.

axiom_mkb

ORVL-004 — Modular Constitutional Knowledge Blocks — parse .axiom specs into typed, HMAC-signed blocks. Actions: register, find, list.

axiom_cas

ORVL-008 — Constitutional Adversarial Sandbox — blue-team detectors over attack payloads, with fix proposals for weak regions. Actions: defend, report.

axiom_crl

ORVL-011 — Constitutional Reinforcement Learning reward — governance scores → a signed scalar reward. Actions: compute, score.

axiom_immune

ORVL-012 — Constitutional Immune System — antibody detectors over a payload: guard-pattern, manifold-distance, HMAC violation, CANNOT_MUTATE, and semantic similarity.

axiom_fusion

ORVL-025 — Fuse an EventToken's modality layers (text / audio / video / physics / governance) into a signed FusedIntent — each layer votes intent signals weighted by confidence.

Research pipeline (1)

axiom_research

9-agent constitutional research pipeline: hypothesis → literature → simulation → critic → safety → ethics → data → experiment → report. Safety and Ethics agents can HALT early on critical risk. Returns per-step signed manifests; uses the active NIM or Anthropic backend.

Install

Prerequisites

git clone https://github.com/Orivael-Dev/axiom.git
cd axiom
pip install -r requirements.txt
export AXIOM_MASTER_KEY=$(python3 -c 'import secrets; print(secrets.token_hex(32))')

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json on macOS, or %APPDATA%\Claude\claude_desktop_config.json on Windows:

{
  "mcpServers": {
    "axiom": {
      "command": "python3",
      "args": ["/absolute/path/to/axiom/axiom_mcp_server.py"],
      "env": {
        "AXIOM_MASTER_KEY": "<your-64-hex-key>"
      }
    }
  }
}

Restart Claude Desktop. The 23 tools appear in the tool menu.

Claude Code

Add to .mcp.json at the root of any project you want governed by Axiom:

{
  "mcpServers": {
    "axiom": {
      "command": "python3",
      "args": ["axiom_mcp_server.py"],
      "env": {
        "AXIOM_MASTER_KEY": "<your-64-hex-key>"
      }
    }
  }
}

Path is relative to the project root once you clone the repo into it (or a sibling directory adjusted with an absolute path).

Cursor

Cursor reads MCP servers from ~/.cursor/mcp.json (global) or .cursor/mcp.json (per-project). Same JSON shape as Claude Code, but prefer an absolute path so Cursor can find the script from any project root:

{
  "mcpServers": {
    "axiom": {
      "command": "python3",
      "args": ["/absolute/path/to/axiom/axiom_mcp_server.py"],
      "env": {
        "AXIOM_MASTER_KEY": "<your-64-hex-key>"
      }
    }
  }
}

Generic stdio

Any MCP client that speaks JSON-RPC 2.0 over stdio works. Spawn the process with AXIOM_MASTER_KEY in the environment and read/write line-delimited JSON-RPC on stdin/stdout:

AXIOM_MASTER_KEY=<hex> python3 axiom_mcp_server.py

Hosted manifest

A machine-readable description of the server lives at orivael-dev.github.io/axiom/mcp.json — 23 tool entries with input schemas, four install snippets, signing metadata. Curl it, grep the right block, paste into your client:

curl -s https://orivael-dev.github.io/axiom/mcp.json | jq .install.claude_code.snippet

The manifest is the same document the README links to.

Signing & verification

Every tool result body carries an hmac_signature field. Verify client-side:

from axiom_signing import derive_key
import hmac, hashlib, json

key = derive_key(b"axiom-mcp-v1")
payload = {k: v for k, v in result.items() if k != "hmac_signature"}
canonical = json.dumps(payload, sort_keys=True, separators=(",", ":")).encode("utf-8")
expected = hmac.new(key, canonical, hashlib.sha256).hexdigest()
assert hmac.compare_digest(result["hmac_signature"], expected)

If the signature verifies, the response was produced by this server and not modified in transit. If it doesn't, treat the response as untrusted — do not act on it.

Troubleshooting