mcpField
P
pranaxis-mcp-gateway

pranaxis-mcp-gateway

Local
github.com/AnaciberAI & Agents0Updated 3w ago

About pranaxis-mcp-gateway

pranaxis-mcp-gateway is a stdio proxy for coordinating resource-consuming MCP calls across agents. It passes reads through and blocks conflicting calls against the same resource version, returning an explanatory verdict and certificate.

Key features

  1. 01Arbitrates calls consuming shared resources
  2. 02Passes read calls through
  3. 03Uses declarative per-server rules
  4. 04Logs verdicts and certificates
  5. 05Supports software or physical verifier

Use cases

  1. 01Prevent agents from consuming the same version twice
  2. 02Coordinate writes through per-server rules
  3. 03Inspect decisions in a JSONL log

MCP support

Sits between an MCP client and server as a stdio proxy. It arbitrates calls matching configured consumption rules and passes unmatched calls through.

Configuration

Local server config is the same for every client. Copy it into your client’s MCP config file.

mcp.json · Universal config
{
"mcpServers": {
"anaciber-pranaxis-mcp-gateway": {
"command": "uvx",
"args": [
"pranaxis-mcp-gateway"
]
}
}
}
Run commanduvx pranaxis-mcp-gateway

Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.

Packages
pypipranaxis-mcp-gateway0.3.0

Tools

No tool list yet

Local servers expose tools only once running on your machine, and the maintainer has not provided a tool list.

Details

From the project README

Pranaxis MCP Gateway

Consistency verdicts for MCP tool calls. A transparent stdio proxy that sits between any MCP client (Claude Code, Cursor, Codex, custom agents) and any MCP server. Reads pass through untouched. Calls that consume a shared resource (launch a run, pay an order, write a file at a given sha, update a row) are arbitrated first: if another agent already holds that resource version, the call never reaches the tool and the agent gets a readable error with a certificate.

It solves the lost update of multi-agent systems: two agents, each inside its own policy, consuming the same thing twice.

agent ──stdio──▶ pranaxis-mcp ──stdio──▶ your MCP server ──▶ tool
                     │
                     └── verdict (software, or the Pranaxis chip)

Install

pip install pranaxis-mcp-gateway      # or: uvx pranaxis-mcp-gateway

Use

Register the gateway in your MCP client instead of the server, with the server as the child command. One gateway per agent, each with its own --agent-id; all gateways on a machine share the version state. Claude Code example, AMD Ross Vivado server:

claude mcp add vivado-mcp --scope project --transport stdio --env VIVADO_PATH=/path/to/vivado -- \
  pranaxis-mcp --agent-id agent-a --rules vivado-ross -- vivado-mcp-server --stdio-bridge

GitHub MCP server:

pranaxis-mcp --agent-id agent-a --rules github-official -- npx -y @modelcontextprotocol/server-github

A denied call returns isError: true with a message like:

DENIED by the Pranaxis consistency verifier.
Resource 'owner/repo/main/README.md' version abc123 is being consumed by another agent
(holder: agent-a; their request: agent-a-3f2c...). Certificate: ...
Your call did NOT reach the tool. Do not retry the same write; read the current state and work from it.

Tool descriptions are annotated so the agent knows writes are arbitrated. In our tests, agents (two different models) did not retry blindly after a denial: they read the state and chose another action.

Rules: what counts as consumption

Rules are data, one JSON file per MCP server (src/pranaxis_gateway/rules/). Each rule names the tool, matches arguments with regexes (named groups become fields), builds the resource name from a template, and optionally takes the version from an argument (e.g. GitHub's previous blob sha). release rules say which responses show a resource complete. A completed resource stays with its holder until the holder makes its next call (it collected the result) or a grace period expires (--grace, default 600 s): nobody wipes someone else's result before they read it.

Bundled: vivado-ross (AMD Ross Vivado MCP server), github-official, postgres-generic. Calls matching no rule pass through and are logged as passthrough. Contributions of rules for other servers are welcome.

Verifiers

  • --verifier stub (default): software reference. Holders shared through the state file; verdicts carry no physical measurement.
  • --verifier fam --fam-endpoint host:port: the Pranaxis physical verifier, a ring-oscillator block on an AMD Zynq UltraScale+ / Kria device that measures the arbitration and returns a certificate with sieve, frequencies, tolerance and timing. The hardware is a separate product (https://pranaxis.eu); this repository contains only the gateway and the HTTP contract it speaks.

Certificates

Every verdict is appended to ross_demo_YYYYMMDD.jsonl (name configurable with --log-file): agent, resource, version, decision, reason, conflicting request, certificate id, measurement data when physical, proxy latency. Read-only calls are logged as passthrough.

Tests

python tests/test_two_agents.py --verifier stub        # two agents, Vivado-like server: P1..P4
python tests/test_rules_generic.py                      # GitHub and SQL rule files

Status and roadmap

0.3: local mode (one proxy per agent, stdio), declarative rules, holder release / grace, jsonl log. Measured with the physical verifier on a ZUBoard 1CG (10/10 preregistered runs) and with two real agents on AMD Ross. Next: central mode (one network gateway for all agents, audit API), embedded mode on Kria K26.

Intellectual property and licence

Code: Apache-2.0. The arbitration procedure and the physical device are covered by Spanish patent applications P202631184 and P202631345 (Arignatxa S.L. as licensee); using this gateway with the software verifier is free; the physical verifier bitstream is not part of this repository. "Pranaxis" is a trademark application (M4406608).

Comments

Markdown supported

New to MCP?

From what the protocol is to installing your first server, these answers cover most first questions.

An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.