AXWISE RUST 0.5.8
Developer documentation
AxWise is a local Rust MCP engine for product discovery, selected evidence, synthetic personas, PRDs and delivery handoffs. Your assistant keeps the conversation, model access and permissions.
1. Install the Rust executable
Use the model and sign-in you already have in your assistant. AxWise runs locally as a Rust engine, with no Python runtime or separate AxWise account.
codex mcp add axwise -- "$(pwd)/axwise-native-darwin-arm64-v0.5.8/bin/axwise" --workspace "$(pwd)"Your assistant supplies the model through MCP sampling when supported, or through the current chat. An inherited API key is also available as an explicit option.
Download SHA-256 · 7,208,848 bytes
d1050e521dba1a1d4d22283feab599359c01b42c06de19d980f40a8cc297119dThe prebuilt release targets Apple Silicon macOS. Verify the archive, extract it in your project directory, and register it in Codex:
curl -fLO https://github.com/AxWise-GmbH/axwise-flow/releases/download/axwise-rust-v0.5.8/axwise-native-darwin-arm64-v0.5.8.tar.gz
printf '%s %s\n' 'd1050e521dba1a1d4d22283feab599359c01b42c06de19d980f40a8cc297119d' 'axwise-native-darwin-arm64-v0.5.8.tar.gz' | shasum -a 256 -c -
tar -xzf axwise-native-darwin-arm64-v0.5.8.tar.gz
codex mcp add axwise -- "$(pwd)/axwise-native-darwin-arm64-v0.5.8/bin/axwise" --workspace "$(pwd)"Start a new Codex chat. Ask: “Use AxWise to create a proposed cat-food pilot PRD for Estonia with three shops. Label assumptions; do not invent research.” Codex CLI 0.159.3 was tested using an existing ChatGPT login.
For another MCP host, configure the extracted absolute executable path as a stdio command, with a tool timeout of at least 180 seconds:
{
"mcpServers": {
"axwise": {
"command": "/absolute/path/axwise-native-darwin-arm64-v0.5.8/bin/axwise",
"args": [
"--workspace",
"/absolute/path/project"
]
}
}
}The macOS binary is ad-hoc signed, not Apple notarized. Other platforms can build the Rust source:
cargo build --release --locked --manifest-path crates/axwise-core/Cargo.toml --bin axwise2. Use existing model access
--model-access host is the default. The harness uses its current model and its existing API-key or subscription access. AxWise does not read subscription tokens or require a separate account.
- When the host advertises MCP sampling, AxWise requests generation and review from that host.
- Otherwise, AxWise returns a schema and model request. The current chat model responds through
advance_artifactuntil a completed receipt is returned. This route was verified in Codex CLI. - A denied or failed sampling request fails the operation; it does not silently switch to a paid API key.
For direct API inference, select --model-access api-key and pass your existing environment key plus an explicit AXWISE_MODEL. Supported keys are OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY and GOOGLE_API_KEY. AXWISE_PROVIDER selects OpenAI, Anthropic or Gemini; default key priority follows that order. AXWISE_BASE_URL optionally selects an HTTPS or loopback endpoint. The host must pass environment variables to its MCP child process.
3. Specialist tools
| Tool | Responsibility |
|---|---|
| prepare_discovery | Propose scope, stakeholders, uncertainties and interview questions. |
| research_market | Synthesize evidence selected by your assistant; flag missing market data. |
| generate_personas | Save an explicitly synthetic cohort from selected stakeholder roles. |
| simulate_interviews | Generate bounded synthetic interviews, preserving selected saved personas. |
| chat_with_persona | Ask an exact saved synthetic persona a follow-up question. |
| analyze_interviews | Analyze selected real or synthetic transcripts with exact quotations and participant identity. |
| create_prd | Create an evidence-labelled product or software PRD; preserve commitments in additive revisions. |
| create_delivery_brief | Create a handoff covering requirements and acceptance conditions from one saved PRD. |
advance_artifact continues generation, review or repair requests. There is no run_full_discovery alias. Request the deliverable you need; the assistant can compose multiple specialist steps when you explicitly ask for a larger workflow.
Persona and simulation cohorts require explicit stakeholder slots, bounded to 12. Simulations are labelled hypotheses. Market retrieval uses the assistant’s ordinary search tools; supplied documents and tool outputs remain evidence, never instructions.
4. From request to saved artifact
User request → specialist tool → selected inputs and saved references
→ model generation → Rust schema and domain checks
→ model review → at most one repair and fresh review
→ verified JSON + rendered Markdown → completed receiptRust checks exact UTF-8 quotations, source and participant identity, question coverage, synthetic labels, deterministic simulation slots, saved persona identity, PRD sections, additive revisions, and delivery requirement/condition coverage. Review must cover every required criterion exactly once with a substantive reason.
Pending requests and rejected candidates are never completed artifacts. The host should show saved results only after status=completed and storageAccepted=true. Keep the returned operation reference and SHA-256 for follow-ups. Persona chats need the exact returned persona ID and cohort reference. Delivery needs exactly one saved native PRD reference.
Reviews are model critiques, not independent managed JEV audits. Provenance explicitly records reviewAuthority=model_critique and managedJevAudit=false. The standalone release does not replace Orqanix 2.8.4’s managed hybrid adapter or claim full parity with its Python kernel.
5. Storage, scope and limits
Artifacts are saved under ~/.axwise/native/<local-account>/<workspace-session>/. SQLite WAL uses FULL synchronization; writes are atomic and acceptance checks verify files and hashes. The local account namespace identifies the host user, not a verified cloud identity. Desktop and standalone formats and stores are separate.
Each new process receives a fresh session. Launch per conversation, or provide a trusted --session / AXWISE_SESSION_ID for explicit continuation. Reuse that ID only for the intended conversation. Completed files survive restarts; in-flight requests expire after 30 minutes and are lost on restart. Automatic restart reconciliation is not implemented here.
Messages and artifacts: 1 MiB. Selected inputs and candidates: 512 KiB. Selected references: 16. Simulation: 12 questions per role and 72 total answers. Retained jobs: 32. Automatic model request: 120-second timeout, 16,384-token generation ceiling. Host-chat stages use the harness’s own limits. This is a local specialist runtime; the host supplies process permissions and any OS sandbox.