# AgentMaurice

> Proprietary agentic infrastructure for building, deploying, and operating governed AI Agents.

## Single prompt
Paste one of these, then follow this file before any other guide:

- FR: Configure AgentMaurice pour cet assistant en suivant https://agentmaurice.ai/llms.txt, puis aide-moi à créer un outil de suivi de mes clients. Réutilise mon installation si elle existe.
- EN: Set up AgentMaurice for this assistant using https://agentmaurice.ai/llms.txt, then help me build a customer-tracking tool. Reuse my existing installation if there is one.

## Decide the surface first
Do not download or mutate until you know which row applies.

1. Local coding agent with a terminal on the requested machine -> follow https://agentmaurice.ai/en/one.md (FR: https://agentmaurice.ai/fr/one.md). Stop before download on Windows, Intel Mac, or if this shell is not the requested machine.
2. Existing Maurice/AgentMaurice install or instance reachable -> reuse identity, data directory, Skill and linked context. Do not create a second install by default. If several targets exist, ask for an explicit choice.
3. Hosted assistant without access to the person's machine -> orient to a qualified remote connector/plugin; guide a human local install and keep the pending request. Never claim a local install completed from a cloud terminal.
4. Container or remote/SSH host -> confirm that environment is the intended install target. If network is forbidden or the surface is unsupported, diagnose and stop; do not bypass network controls.

First proof lot: Codex and Claude Code with a local terminal. Other surfaces must be oriented honestly; do not announce full support without a matching recipe.

## Load details only as needed
- Compact router (this file): https://agentmaurice.ai/llms.txt
- Local install / reuse / pairing: https://agentmaurice.ai/en/one.md | https://agentmaurice.ai/fr/one.md
- Human page (same path): https://agentmaurice.ai/en/for-agents | https://agentmaurice.ai/fr/for-agents
- Versioned decision contract: https://agentmaurice.ai/agent-guide/v1.json
- Canonical Skill: https://agentmaurice.ai/skill/agentmaurice/SKILL.md
- Skill manifest: https://agentmaurice.ai/skill/manifest.json

## AgentMaurice One: local setup for coding agents
AgentMaurice One is the free, proprietary local runtime for personal and professional use. Build Agents, Modules and Applications locally; use a paid AgentMaurice server offer when sharing or scaling the runtime.

Install Maurice from https://agentmaurice.ai/en/one.md. The script downloads the archive from get.agentmaurice.app and checks its SHA-256. The stable CLI manifest still lists v2.1.29; a CLI download alone does not install Maurice.
Availability: public_alpha; checked 2026-09-29; installable_from_this_guide=true.
Observed CLI manifest (not a Maurice installer): https://get.agentmaurice.app/products/mauricecli/latest.json
Install procedure: https://agentmaurice.ai/en/one.md; French: https://agentmaurice.ai/fr/one.md.
Full Maurice guide: https://agentmaurice.ai/en/for-agents#one; JSON: https://agentmaurice.ai/agent-guide/v1.json (one.en or one.fr).
WebMCP: call get_one_guide with {}. It only returns guidance; it does not install or start anything.

Follow the procedure at https://agentmaurice.ai/en/one.md. The script selects the newest public archive for this machine and accepts only the published flags. Do not guess another installer, download URL, or flag.
macOS arm64 is signed. Linux amd64 and arm64 are checksummed. Mac Intel and Windows have no public archive yet.

### For an already provisioned Maurice installation
1. Inspect the installed command contract. Continue only if start, status, doctor and stop describe the local Maurice lifecycle.
   Command: maurice docs --full --format json
2. Start explicitly. For a fresh start require ready=true and ready_checked=true. If already_running=true, verify the existing instance with status and doctor.
   Command: maurice start --json
3. Verify the process is running.
   Command: maurice status --json
4. Require running, health_checked, health_ok, storage_checked and storage_ok. The local Doctor does not verify LLM access or MCP execution.
   Command: maurice doctor --json

Then load the canonical Skill. Inspect the CLI context and use only the authentication and Agent scope supplied by this installation. Inspect maurice tools list before concluding that a runtime tool is unavailable. Configure the required LLM or MCP credentials through the instance, then validate one real operation before claiming the Application works. If local authentication or a required capability is missing, report that prerequisite instead of inventing a bootstrap.
Stop only the local instance you started when it is no longer needed. If you use --data-dir, keep the same value for every lifecycle command. Do not publish the local runtime on a public interface.
Shutdown command: maurice stop --json
Doctor still needs --data-dir, and a workflow call still needs --runtime-url http://127.0.0.1:5011, until the connected context keeps those values. Windows and Mac Intel have no public archive. Do not present the stable CLI manifest as a Maurice installer.

## Canonical object model
- Agent: A deployed, operable AI product resource managed by AgentMaurice. An Agent is produced from one Agent Spec and owns managed Workflows and MiniApps.
- Agent Spec: The versioned, Git-native desired state for one Agent. It is checked, planned, approved when required, applied, and verified.
- Workflow: An executable business process managed by an Agent Spec. Workflows orchestrate models, MCP tools, code, data, and human gates.
- MiniApp: An interactive runtime surface managed by an Agent Spec. A MiniApp presents guided inputs or results while side effects stay in Workflows.
- Module: A versioned executable package contributing Workflows, MiniApps, schemas, assets, and docs. A Module is installed into an Application; it is not a Skill or an Agent Spec.
- Application: An organizational runtime surface composed of one or more Modules and Agents. Applications are composed with maurice app; Agent Specs remain in consuming Agent projects.

Never confuse these objects: a Module is an executable package installed into an Application; an Agent Spec is the Git-native desired state of one Agent; a Skill is instruction-only.

## Canonical rails
- Agent: connect -> init|pull -> edit -> commit -> spec deploy -> policy authorization or human approval -> apply -> verify.
- Application: module validate/test -> app init/add -> app plan -> human approval when required -> app apply -> app status.
- Module: module init -> module validate -> module test -> install in a test Application -> publish through the governed rail.
- Existing runtime: use External Inception; confirm scopes, run the compact Doctor, then read capabilities before resolving runtime tools.

## Safety
- Public website tools are advisory and read-only. They never connect to an instance.
- Never provide raw secrets, bearer tokens, single-use bootstrap URLs, organization identifiers, or sk_maurice_* values to website tools.
- Use placeholders in commands. Staging and production require an explicit authenticated human approval.
- Authenticated operations stay on MauriceCLI, AgentMaurice OS, or External Inception.

## For coding agents without WebMCP
Read the versioned JSON contract, reproduce its deterministic Doctor rules locally, then load the canonical Skill. WebMCP is a browser enhancement, not an IDE MCP connector. Change rails to MauriceCLI or instance-scoped External Inception only when authenticated work is required.