Aethis

aethis-mcp

MCP server for the Aethis decision engine. Compile legislation, policy, contracts, and regulation into deterministic logic — same input, same answer, every time, with a full audit trail.

Install · Skills · Quick start · Tools · Setup · Authoring · DSL · Troubleshooting

Install

Authoring is in private beta. Decision tools (aethis_decide, aethis_schema, aethis_explain, aethis_next_question) are public — no key required. Authoring tools (rule generation, test refinement, publishing) require an invite. Request access at aethis.ai/developer-access.

Recommended — one command via aethis-cli:

uv tool install aethis-cli
aethis mcp install --target all

Wires the server into claude-code, cursor, claude-desktop, or windsurf. Idempotent. Restart your editor to pick up the change. Re-run after aethis account generate rotates a key. Full options: aethis mcp install --help.

Manual install:

claude mcp add aethis -- npx -y aethis-mcp

For Cursor / Claude Desktop / Windsurf manual config, see Setup.

Onboarding an AI coding agent end-to-end? See docs.aethis.ai/agents/onboarding — install + verify + auth + workflow patterns in one page.

Skills

After the MCP server is installed, add reusable agent workflows with aethis-skills:

npx skills add Aethis-ai/aethis-skills

The skills package provides workflows for policy-to-ruleset authoring, test/refine/publish loops, decisions with trace, and regression comparison. It calls the MCP tools in this package; it does not replace the MCP server.

Quick start

aethis_decide({
  ruleset_id: "aethis/spacecraft-crew-certification",
  field_values: { "space.crew.species": "Vogon" },
  include_trace: true
})

{
  "decision": "not_eligible",
  "fields_provided": 1,
  "fields_evaluated": 11,
  "trace": {
    "species_check": "FAIL — species is 'Vogon' (disqualifying, Section 3)"
  }
}

Public rulesets work without a key. Browse: aethis_discover_rulesets({}) or docs.aethis.ai.

Engine determinism + accuracy benchmarks: Aethis-ai/confidently-wrong-benchmark.

Tools

35 tools across six groups.

aethis_graph is public for a public showcase ruleset (ruleset_id) and tenant-scoped for a rulebook (rulebook_id) — it returns the ruleset-map graph ({nodes, edges, sections, stats}, each node's display.sentence/display.routes/display.expr) plus a ready-to-render mermaid diagram string. Pass include_graph_overlay: true to aethis_decide to get that same graph back with a specific decision's per-criterion status (satisfied/not_satisfied/pending) stamped onto it (graph_overlay in the response) — a "you are here" map for those inputs.

aethis_create_rulebook / aethis_update_rulebook manage a Rulebook's identity (name/domain/slug/description) and robot_hints — beat-keyed natural-language guidance for the conversational agent. Active beats: general_context, preamble, session_start, postamble, session_end, stuck. Reserved (accepted, not yet acted on): persona, conversational_style, section_transition. Composition (bridging rulesets via outcome_logic) is a separate, larger surface not covered by these two tools yet.

Workflows

Evaluate eligibility (2 calls):

aethis_schema(ruleset_id)          → fields needed
aethis_decide(ruleset_id, fields)  → eligible / not_eligible / undetermined

Pass include_trace: true for the per-criterion evaluation trail. Pass include_explanation: true for human-readable rule descriptions.

aethis_decide accepts either ruleset_id (single ruleset, may be public) or rulebook_id (composed multi-ruleset rulebook) — the two are mutually exclusive. Rulebook decide always requires an API key (AETHIS_API_KEY); anonymous callers get HTTP 401. aethis_graph follows the same ruleset_id/rulebook_id split for the underlying map.

Conversational eligibility (next-question routing):

aethis_next_question(ruleset_id, field_values)

Returns the most informative remaining question and the optimal_path of remaining questions. Call again after each answer; the engine recomputes from the updated state. Stops when a decision is reachable.

Authoring (private beta): see Authoring.

Prompts

Setup

Decision tools work with no key. For invited authoring access, run aethis login, then install with aethis mcp install --target <client>. The installer references a saved profile so the host configuration does not contain the API key.

Claude Code

# Decision tools only
claude mcp add aethis -- npx -y aethis-mcp

# With authoring access
claude mcp add aethis -e AETHIS_PROFILE=default -e XDG_CONFIG_HOME=/absolute/path/to/config -- npx -y aethis-mcp

Claude Desktop

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

{
  "mcpServers": {
    "aethis": {
      "command": "npx",
      "args": ["-y", "aethis-mcp"]
    }
  }
}

For authoring, add "env": { "AETHIS_PROFILE": "default", "XDG_CONFIG_HOME": "/absolute/path/to/config" }. Use your saved profile name and the absolute directory containing aethis/credentials (normally your home directory’s .config).

Cursor / Windsurf

Add to ~/.cursor/mcp.json or ~/.codeium/windsurf/mcp_config.json (same JSON shape).

Keys

• AETHIS_PROFILE — non-secret saved profile name. It pins the account and endpoint used by this registration, even if the CLI’s active_profile later changes.
• XDG_CONFIG_HOME — absolute config directory containing aethis/credentials. Relative values and credentials symlinks escaping your home/config directory are refused; credential files must have no group/other permission bits (normally 0600).
• AETHIS_API_KEY — optional deliberate process-environment override for the platform key. Prefer saved-profile references when installing; avoid putting raw keys in host config or command arguments. The host must securely supply the process environment; it may not inherit your shell environment.
• AETHIS_ANTHROPIC_KEY_ENV — the name of the env var holding the Anthropic key Aethis authoring tools may use (e.g. AETHIS_ANTHROPIC_KEY). The server never reads a provider key from the environment unless you set this. See Passing your Anthropic key safely.
• Rotate via aethis account generate + aethis account revoke <key_id>. Mint one key per machine for surgical revocation.

Credential precedence and restart behavior

MCP parses the CLI credentials file as YAML. AETHIS_PROFILE selects a named profile; otherwise the file’s active_profile (or default) selects it. The profile supplies both its API key and base_url, with https://api.aethis.ai as the default endpoint. With an explicit AETHIS_PROFILE, AETHIS_API_KEY and AETHIS_BASE_URL deliberately override their respective values. Without an explicit profile selector, AETHIS_API_KEY uses AETHIS_BASE_URL or the default endpoint, ignoring the implicitly active profile (including implicit anonymous). Missing or malformed explicitly selected profiles fail visibly, including when environment overrides are present. Explicit AETHIS_PROFILE=anonymous always stays unsigned. After anonymous setup, run aethis login and install again to reference the saved authoring profile, then restart the host.

A saved profile outranks old macOS Keychain entries. If no profile is configured and no explicit name is selected, MCP can use the legacy default keychain entry, then the older flat credentials.yaml file. Flat api_key/base_url files at aethis/credentials remain supported. A configured profile awaiting login stays unsigned instead of borrowing another stored key.

The server keeps its authenticated startup key and endpoint paired until restart. If it started without a key, an authenticated tool can pick up a later login for the same endpoint. A changed endpoint causes a visible refusal: restart the MCP host to load the new pair. Startup stderr reports only the credential source, never key values.

Passing your Anthropic key safely

Authoring tools (aethis_generate_and_test, aethis_refine, aethis_discover_fields, aethis_refine_fields, aethis_discover_sections, aethis_refine_sections) need an Anthropic API key per call. Three accepted forms — listed in preferred order:

The server sends a provider key to Aethis only when you configured one for Aethis. It never picks up ANTHROPIC_API_KEY (or any other variable) from your environment on its own, and an env-var name supplied in a tool call by the host model is refused unless it is the one you configured. Only Anthropic keys (sk-ant-…) are accepted; anything else is refused locally and not sent.

1. AETHIS_ANTHROPIC_KEY_ENV (recommended). In the MCP server config, put the key in a dedicated env var and name that var in AETHIS_ANTHROPIC_KEY_ENV. Tools then use it automatically. The raw value never appears in the tool call payload, so it does not land in the MCP host's session transcript on disk.

   // claude_desktop_config.json
   {
     "mcpServers": {
       "aethis": {
         "command": "npx",
         "args": ["aethis-mcp"],
         "env": {
           "AETHIS_PROFILE": "default",
           "XDG_CONFIG_HOME": "/absolute/path/to/config",
           "AETHIS_ANTHROPIC_KEY_ENV": "AETHIS_ANTHROPIC_KEY",
           "AETHIS_ANTHROPIC_KEY": "sk-ant-..."   // never echoed back to the LLM
         }
       }
     }
   }
   

   aethis_generate_and_test({ project_id })   // uses the configured key
   

2. anthropic_key_keychain (macOS). A keychain reference — either "account" (service defaults to aethis-anthropic-key) or "service:account". Store the key once with security add-generic-password -U -s aethis-anthropic-key -a my-anthropic -w 'sk-ant-...', then call:

   aethis_generate_and_test({ project_id, anthropic_key_keychain: "my-anthropic" })
   

3. anthropic_key (deprecated). Pass the raw key as a tool argument. Accepted for backwards compatibility, but the raw value is written verbatim to the host's session transcript JSONL on disk. If a key was ever passed this way, rotate it before relying on the safer forms.

Authoring (private beta)

Authoring requires an invite. Request access. Decision tools (above) are public.

Three-phase workflow. Phases 1–2 are for multi-section domains; skip them for single-section rules and go straight to Phase 3.

Phase 1 — Section discovery

aethis_discover_sections({ domain, sources: [{ name, content }, ...] })
aethis_validate_sections({ domain, expected_sections, discovered_sections })
aethis_refine_sections({ domain, feedback, sources })

Phase 2 — Field vocabulary

aethis_set_field_spec({
  project_id,
  expected_fields: [{ key, sort, enum_values?, notes?: [{ note_text, source?, metadata? }] }, ...]
})
aethis_discover_fields({ project_id })           // auto-validates against the spec if set
aethis_refine_fields({ project_id, feedback })
aethis_validate_fields({ project_id, expected_fields })

Phase 3 — Generate, test, publish

aethis_create_ruleset({
  name, section_id, domain?, source_text,
  test_cases: [{ name, field_values, expected_outcome, expectations? }, ...],
  contract_version?: 1,
  expected_review_bindings?: { field_id: { token: true | false | null } }
})
aethis_generate_and_test({ project_id })
aethis_refine({ project_id, feedback })          // iterate until tests pass
aethis_publish({ project_id })                   // refuses if tests fail; returns ruleset_id on success

When a test carries expectations, set contract_version: 1. The optional binding catalogue is generic authoring metadata: omit it when no binding assertion is needed, or pass {} to assert that no review bindings exist. The server must confirm the complete stored contract before generation begins. aethis_create_ruleset creates a project; it does not append or replace tests on an existing project. Use aethis_set_tests with the complete version-1 contract for a later replacement, then generate or refine that same project. Legacy updates cannot discard stored assertions. Object keys named __proto__ are rejected in field-value and binding maps rather than silently discarded.

If generation polling times out, call aethis_generation_status({ project_id }) before retrying: use its telemetry_availability, server-authoritative worker_lifecycle, and retry_readiness, and retry only when readiness is ready. An old heartbeat alone is not proof that the worker died. Call aethis_cancel_generation({ project_id, job_id, confirm_job_id }) only after showing the observed job_id and receiving explicit confirmation to abandon that active run. It releases the project's job ownership, but worker shutdown may be cooperative rather than immediate; inspect the returned detail. It is a destructive, API-key-protected mutation. The response distinguishes a new cancelled transition from the idempotent already_cancelled result.

Guidance

Targeted hints without regenerating, plus cross-section principles for a domain:

aethis_add_guidance({ project_id, guidance_text, process_type })
aethis_list_guidance({ project_id })

aethis_add_domain_guidance({ domain, guidance_text, process_type, notes? })
aethis_list_domain_guidance({ domain })

process_type is rule_generation (default) or field_extraction.

Diagnose a failing test

aethis_explain_failure({
  ruleset_id, field_values, expected_outcome, test_name
})
// Returns criterion statuses, the failing rule, and a targeted fix hint.

[!IMPORTANT] Tests are the publish gate. aethis_publish refuses to publish a ruleset with a failing test. SMEs write the tests; the LLM generates the rules from source text + guidance; the platform refuses to ship rules that don't satisfy the tests. Better tests = faster convergence.

[!IMPORTANT] Anthropic key required for authoring. Configure AETHIS_ANTHROPIC_KEY_ENV or use anthropic_key_keychain (macOS keychain ref) rather than the raw anthropic_key argument — see Passing your Anthropic key safely. Used per-request, never stored server-side; the raw form, however, lands in the MCP host's session transcript on disk.

[!IMPORTANT] DATE fields use integer ordinals (date.toordinal()), not ISO strings. 2025-04-13 = 739354. Quick conversion: python3 -c "from datetime import date; print(date(2025,4,13).toordinal())".

Field types

Operators

Helpers

• days_between(date_a, date_b) → Int
• years_between(date_a, date_b) → Int — completed whole years between the two dates (leap-correct). Use this for age from a date-of-birth field; never derive age as days_between(...) / 365.
• min(a, b, ...), max(a, b, ...) → Int
• Constant arithmetic folded at authoring time (5 * 365 → 1825)

Not supported

• Division between runtime field values
• Weighted scoring or probabilistic outcomes
• Lists as field values (use pre-aggregated Int / Bool)
• More than 3 outcome tiers (eligible / not_eligible / undetermined)

Troubleshooting

Related

• aethis-cli — Python CLI; file-based authoring with YAML test cases
• aethis-examples — runnable rulesets (spacecraft, construction-CAR, consumer credit) and benchmark scenarios
• confidently-wrong-benchmark — paper, 225-scenario benchmark, LegalBench harness

Development

git clone https://github.com/Aethis-ai/aethis-mcp.git
cd aethis-mcp && npm install && npm test && npm run build

License

MIT