grimoire-kit 3.37.0

The governed operating system for AI agents in your IDE — verify, don't trust.

_{English — version française (complète)}

Quick Start • Governed Standard • Architecture • Features • Full guide • Changelog

Why Grimoire Kit?

AI coding assistants run on trust: you hope the agent kept its guardrails, remembered the context, and produced evidence for what it claims. Grimoire Kit replaces that hope with verifiable governance:

• A governed agentic standard — 36 declarative patterns (security, memory, context, observability, quality…) verified fail-closed by grimoire standard verify / audit / score / gate check. If a required proof is missing, the gate fails.
• Agent teams with personas and semantic memory — specialized agents, persistent memory across sessions, delivery contracts, quality automation.
• One config, every assistant — grimoire init generates portable entrypoints (CLAUDE.md, AGENTS.md, GEMINI.md, .cursorrules) plus a .mcp.json, so the same project works with Copilot, Claude Code, Codex, Gemini CLI and Cursor.
• Native on each host, not lowest-common-denominator — one host-neutral description of the project is rendered per host: sub-agents with their own context window and tool boundary, on-demand skills, slash commands, and lifecycle hooks that actually refuse a destructive call, or the closure of a task whose evidence gates are red. What a host cannot execute is reported as an explicit degradation, never dropped in silence. See grimoire host --help and docs/hosts.md.

Quick Start

# Install (uv recommended)
uv tool install grimoire-kit

# Or with pipx
pipx install grimoire-kit

# One command: init + identity + governed standard + health check
grimoire up my-project --archetype web-app

# Existing project (idempotent — re-running only fills the gaps)
cd your-project/
grimoire up . --name "My Project"

# Detect and enroll existing projects on your machine
grimoire cockpit scan ~/dev

# Health check + repair (regenerates missing wrappers and .mcp.json)
grimoire doctor . --fix

# Local multi-project dashboard (background + browser)
grimoire cockpit

grimoire init . --name "My Project"   # scaffold: context, agents, memory, entrypoints
grimoire setup                        # propagate identity into instructions
grimoire standard init . --needs ...  # governed standard artifacts
grimoire doctor                       # health check

Multi-assistant — grimoire init writes portable assistant entrypoints pointing to a single canonical instruction file, and registers the Grimoire MCP server in .mcp.json (OS-neutral, via grimoire-mcp). No per-assistant duplication.

The governed agentic standard

Declare what your project needs; Grimoire resolves a maturity profile (starter → controlled → orchestrated → governed → production) and activates governed patterns — 36 in the catalog. Each pattern scaffolds a declarative artifact (_grimoire/standard/*.yaml) and a fail-closed check:

grimoire standard needs                        # needs grouped by tier, with footprint
grimoire standard plan --needs solo-prototyping  # preview without writing anything
grimoire standard init . --needs solo-prototyping # minimal recommended start
grimoire standard init . --interactive         # guided setup

grimoire standard verify       # artifacts present + conform to the profile
grimoire standard audit        # compliance report + remediation actions
grimoire standard score        # multi-dimensional compliance score
grimoire standard gate check   # CI gate: fails when required evidence is missing

Example controls: tool blast-radius limits, controller/agent privilege separation, prompt-injection firewall, decision council, memory integrity, LLM cost registry + reliability SLO, guardrail contracts (input/output/tool/model), visual evidence, workspace isolation, per-environment policies.

Reference: governed controls (36 patterns) · standard integration · install by needs

Architecture

Three layers, one wheel (pip install grimoire-kit):

Details in ARCHITECTURE.md.

Features

Experimental R&D features (session branching, agent darwinism, stigmergy, dream mode…) are documented separately in docs/rnd.md — functional and tested, but exploratory surface.

CLI, SDK and MCP

grimoire init | doctor | status | up | validate     # project lifecycle
grimoire add <agent> | remove <agent> | registry    # agents & archetypes
grimoire standard …                                 # governed standard (see above)
grimoire cockpit …                                  # local dashboard daemon
grimoire memory …                                   # governed memory operations

from pathlib import Path
from grimoire.tools import PreflightCheck

report = PreflightCheck(Path(".")).run()
print(report.status)  # GO / GO-WITH-WARNINGS / NO-GO

MCP server (grimoire-mcp, stdio) exposes 12 tools: project_context, status, agent_list, config, harmony_check, memory_store, memory_search, add_agent, standard_verify, standard_audit, standard_score, standard_gate — agents can query and enforce the governed standard directly.

Quality

• 5 990+ tests (pytest), coverage gate ≥ 70 % (Codecov)
• mypy strict on the SDK, ruff lint + format
• CI matrix: Ubuntu / Windows / macOS × Python 3.12 / 3.13, CodeQL, dependency audit with governed, time-boxed waivers
• Releases: SBOM (CycloneDX), smoke tests, PyPI trusted publishing (OIDC)

Documentation

Getting started · Concepts · SDK guide · CLI reference · MCP integration · FAQ — full docs (French) on the documentation site.

Contributing & License

Contributions welcome — see CONTRIBUTING.md. MIT — see LICENSE.