axon-emcp

The official ℰMCP (Epistemic Model Context Protocol) server for AXON — a stdio JSON-RPC 2.0 Model Context Protocol server that exposes the AXON language to AI coding agents (Claude Code, Codex, Cursor, Continue, Cline, …).

docs.rs

What it does

Any MCP-compatible coding agent that launches axon-emcp as a subprocess gets:

• 6 tools — axon.primitives, axon.primitive_doc, axon.check, axon.parse, axon.compose, axon.examples. Live language reference + structured validation through the same axon-frontend lexer/parser/type-checker the axon CLI uses (byte-identical diagnostics), plus a curated drift-gated example corpus organised by language idea.
• 14+ resources under axon:// — full primitive references (primitives/{name}), grammar maps (grammar/{top_level|composition|ebnf}), composition logic (logic/{flow_composition|session_duality}), and per-framework compliance maps (compliance/{hipaa|gdpr|pci_dss|sox|soc2|fedramp|gxp|fisma|nist_800_53}).
• 3 prompts — flow_design, shield_design, session_design. Host-surfaced design recipes that drive the agent through a structured workflow (axon.compose → axon.primitive_doc → axon.check).

The result: a developer talks to their AI coding agent in natural language; the agent writes AXON that passes axon check on first try, with the right compliance shields by construction.

Install

cargo install axon-emcp

The installed binary is fully self-contained — the knowledge corpus (99 primitive docs + 34 templates + 30 idiomatic examples + grammar/logic/compliance references) is baked into the executable via include_dir! at compile time. No share/ directory, no env vars, no post-install steps.

Listed on the official MCP Server Registry

The server is registered at registry.modelcontextprotocol.io under the canonical namespace com.ricardovelit/axon-emcp. Tools that consume the registry (Claude Desktop's connector picker, Cursor's MCP browser, the mcp-publisher CLI's discovery surface, future clients that index registry.modelcontextprotocol.io/v0/servers) list the server automatically. Adopters that prefer the registry- driven install flow can run:

# Inspect the canonical metadata:
curl -s 'https://registry.modelcontextprotocol.io/v0/servers?search=axon-emcp' | jq .

# Or use the publisher CLI for inspection:
mcp-publisher status com.ricardovelit/axon-emcp

The crate itself stays on crates.io (the registry currently lists npm | pypi | oci | nuget | mcpb as package-source types; crates.io is registry-tracked via the websiteUrl field until it lands in the package-type enum).

axon-emcp --help

Configure your agent

Point the agent's MCP config at the installed binary:

// Claude Code:    ~/.config/claude-code/mcp.json
// Cursor:         ~/.cursor/mcp.json
// Codex:          ~/.codex/mcp.json
// Continue:       ~/.continue/config.json (under "mcp.servers")
{
  "mcpServers": {
    "axon": {
      "command": "axon-emcp"
    }
  }
}

Restart the agent. The 6 tools, 14+ resources, and 3 prompts now appear in the agent's surface — the agent will quote them when you ask it to write AXON.

Closed-domain composition

You: write me a healthcare flow that handles PHI.

Agent: [calls axon.compose(intent: "healthcare flow PHI")]
       ✓ Scaffold returned (axon_check_verdict: well-formed)
       ✓ Domain: healthcare (matched: patient, phi, hipaa, …)
       ✓ Compliance applied: HIPAA, GDPR, GxP, SOC2
       ✓ Primitives used: type, persona, shield, flow, axonendpoint

       Here is your scaffold: …

The classifier picks from a closed 33-domain catalogue spanning verticals (healthcare, banking, government, legal, legaltech, fintech, pharmatech, medic_research), agent patterns (chat_research, chat_tools, chat_skills, whatsapp, voice, dev, sales_consultive, sales_widget), and application patterns (workflow_automation, business_intelligence, corporate_integration, self_learning, document_analysis, ticket_triage, content_moderation, knowledge_extraction, compliance_monitoring, recruitment, education, financial_advisor, data_pipeline). Every template proven to compile through the live axon-frontend pipeline.

Focused examples library

Where axon.compose gives you a full-app scaffold by domain, axon.examples returns minimal complete programs (~20–60 LOC) organised by idea:

You: show me how to use weave correctly.

Agent: [calls axon.examples(primitive: "weave")]
       ✓ Match: weave_braid — Weave braid — composing multiple sub-derivations
       ✓ Topic: composition
       ✓ Primitives: persona, flow, step, weave
       ✓ Source: (verified to compile through axon-frontend)

       Here is the canonical pattern: …

Filter on three independent axes (combine freely):

• name: — single-example resolution by slug, returns full .axon source the agent can paste / axon.check directly.
• topic: — closed 10-entry taxonomy (composition, session_types, shields, effects, streaming, data, agents, endpoints, memory, validation).
• primitive: — every example that exercises a given primitive name.

Every example is drift-gated through the same axon-frontend pipeline axon.check uses, so what the agent receives is guaranteed to compile.

Contributor surface

Beyond the MCP server mode (no arguments), axon-emcp ships two contributor-facing subcommands:

# Scaffold a new primitive doc with frontmatter pre-populated from the
# canonical PRIMITIVE_REGISTRY.
axon-emcp scaffold-primitive <name>

# Aggregate a JSONL telemetry log into a structured snapshot.
axon-emcp telemetry summarize <file>

Telemetry (privacy-first)

Opt-in only — no network egress without an explicit env var:

Seven privacy invariants are enforced by construction:

1. AXON source content is never recorded.
2. axon.compose intent strings are never recorded.
3. Tool error messages are never recorded.
4. No remote egress without explicit opt-in.
5. Deployment ID is operator-supplied (default: empty).
6. OTLP exporter only spawns when AXON_EMCP_OTLP_ENDPOINT is set — no socket, no DNS, no task otherwise.
7. The OTLP wire payload is a pure function of the in-process snapshot — every value crossing the wire came through one of the closed-catalog record_* recorders.

OTLP/gRPC exporter

Set AXON_EMCP_OTLP_ENDPOINT and the server pushes a metrics snapshot every 60s to your collector:

AXON_EMCP_OTLP_ENDPOINT=http://localhost:4317 axon-emcp
# Or against Honeycomb / Grafana Cloud / Datadog / Lightstep:
AXON_EMCP_OTLP_ENDPOINT=https://api.honeycomb.io:443 \
AXON_EMCP_OTLP_HEADERS=x-honeycomb-team=YOUR_KEY,x-honeycomb-dataset=axon \
  axon-emcp

The OTLP payload carries 18 metric families keyed by service.name=axon-emcp, service.version, deployment.environment (from AXON_EMCP_DEPLOYMENT_ID):

• axon_emcp.tool.{calls,errors} (Sum, tool= label)
• axon_emcp.tool.duration.{p50,p95,p99}_us (Gauge, tool= label)
• axon_emcp.resource.reads (Sum, uri_family= label)
• axon_emcp.prompt.{calls,missing_required_arg} (Sum, prompt= label)
• axon_emcp.compose.{total,overrides,by_domain} (Sum, domain= label)
• axon_emcp.check.{pass,fail}_by_stage (Sum, stage= label)
• axon_emcp.examples.{total,by_name,empty_responses,by_topic,by_primitive} (Sum, labelled)

The JSONL sink + OTLP exporter compose freely — file-based pipelines (Vector / Fluent Bit / otel-collector filelog receiver) and the direct gRPC push both consume the same in-process snapshot.

License

MIT. See the repository for the full sources, the knowledge corpus under src/knowledge/, and the contributor guide.