Getting Started — Grimoire Kit v3¶
English — version française
This guide takes you from installation to your first Grimoire project.
Requirements¶
- Python 3.12+
- pip, pipx (recommended) or uv
- An editor with Copilot / LLM support (VS Code recommended)
Installation¶
# Via pipx (recommended — automatic isolation)
pipx install grimoire-kit
# Via pip inside a venv
python3 -m venv .venv && source .venv/bin/activate
pip install grimoire-kit
# With MCP (Model Context Protocol) support
pip install grimoire-kit[mcp]
# All extras (MCP + Qdrant + Ollama)
pip install grimoire-kit[all]
Note: On Ubuntu/Debian,
pip installoutside a venv is blocked (PEP 668). Usepipxor a venv.
Check the installation:
Create a project¶
In one command (recommended)¶
grimoire up runs the whole journey: init (express mode), identity
propagation, governed agentic standard, then a health check. The command is
idempotent — run again on an existing project, it only fills the gaps.
# New project
grimoire up mon-projet --archetype web-app --name "Mon Projet" --user "Alice"
# Existing project
cd votre-projet/
grimoire up .
# Full interactive wizard, or without the governed standard
grimoire up . --interactive
grimoire up . --no-standard
Available archetypes: minimal, web-app, infra-ops, platform-engineering,
agentic-standard, creative-studio, fix-loop.
Step by step (what up chains together)¶
grimoire init mon-projet --archetype web-app # scaffold only
cd mon-projet
grimoire setup # identity propagation
grimoire standard init . --needs solo-prototyping
grimoire doctor
Detect your existing projects¶
To list the projects on a machine and enrol them in the cockpit:
grimoire cockpit scan ~/dev # recursive crawl (depth 4 by default)
grimoire cockpit scan ~/dev --yes # enrols every Grimoire project found
Git repositories that are not initialised are listed with the suggestion
grimoire up <path>.
Generated structure¶
mon-projet/
├── project-context.yaml # Centralised configuration
├── _grimoire/
│ ├── _config/
│ │ ├── agents/ # Installed agent references
│ │ ├── manifest.yaml # Project manifest
│ │ └── custom/ # Local overrides
│ ├── _memory/
│ │ ├── shared-context.md # Shared context
│ │ ├── decisions-log.yaml # Decision log
│ │ └── learnings.yaml # Learnings
│ └── core/
│ ├── agents/ # Deployed agents
│ └── workflows/ # Active workflows
└── .github/
└── copilot-instructions.md # VS Code Copilot instructions
CLI commands¶
| Command | Description |
|---|---|
grimoire up [path] |
Full journey: init + setup + standard + doctor (idempotent) |
grimoire init <path> |
Initialise a project (scaffold only) |
grimoire setup |
Synchronise user configuration |
grimoire setup --check |
Audit the synchronisation (CI-friendly) |
grimoire doctor [--fix] |
Check project and environment health; --fix regenerates missing wrappers and .mcp.json |
grimoire cockpit scan <root> |
Detect and enrol existing projects |
grimoire blueprint <cmd> |
Blueprints: new, validate, compile |
grimoire status |
Show project state |
grimoire add <agent> |
Add an agent |
grimoire remove <agent> |
Remove an agent |
grimoire validate |
Validate project-context.yaml |
grimoire check |
Lint + validate + doctor in one pass |
grimoire standard <cmd> |
Governed agentic standard (needs, init, verify, audit, score, gate) |
grimoire merge <source> |
Merge Grimoire files |
grimoire merge --undo |
Undo the last merge |
grimoire upgrade |
Migrate a project from v2 to v3 |
grimoire registry list |
List available agents |
grimoire registry search <q> |
Search for an agent |
Configure your identity¶
After grimoire init, configure your name and language. The setup command
propagates these values into every configuration file of the project (Copilot
instructions).
# Synchronise from project-context.yaml
grimoire setup
# Or specify directly
grimoire setup --user "Alice" --lang "English" --skill-level intermediate
# Check the synchronisation (useful in CI)
grimoire setup --check
Source of truth: project-context.yaml (user section). The setup
command propagates the values to:
.github/copilot-instructions.md— instructions injected into Copilot Chat
Check your project¶
Expected output (excerpt):
OK project-context.yaml found
OK Config valid — project: Demo
OK _grimoire/ present
OK Archetype configured: minimal
OK 7 VS Code agent wrapper(s) in .github/agents/
OK uv available (/home/user/.local/bin/uv)
OK docker daemon reachable (server 29.6.1)
OK Qdrant reachable at http://localhost:6333
OK Ollama reachable at http://localhost:11434
OK .mcp.json server 'grimoire' resolves (grimoire-mcp)
17/17 checks passed
Environment checks (uv, docker, Qdrant, Ollama) are optional: a warning shows
the exact remediation command without blocking. A broken reference in
.mcp.json, on the other hand, is an error.
To repair a project (missing agent wrappers or .mcp.json):
Adopt the governed agentic standard¶
Beyond scaffolding, Grimoire provides an agentic standard: a project need
maps to a profile (starter → controlled → orchestrated → governed →
production) which activates verifiable governed patterns (36 in the
catalogue).
# Choose by need (start small)
grimoire standard needs
grimoire standard init . --needs solo-prototyping
# Verify / audit / score / gate compliance (fail-closed)
grimoire standard verify
grimoire standard audit
grimoire standard score
grimoire standard gate
Control reference: Governed controls · integration: Agentic standard · install by needs: Install by needs.
Multi-assistant portability¶
grimoire init generates portable entrypoints — CLAUDE.md, AGENTS.md,
GEMINI.md, .cursorrules (pointing to .github/copilot-instructions.md) and
an OS-neutral .mcp.json — so the project works with Copilot, Claude Code,
Codex, Gemini CLI and Cursor without manual configuration.
Create a blueprint¶
A blueprint describes an agent pipeline (nodes, edges, contracts) that compiles into a mission pack. The CLI covers the whole cycle:
grimoire blueprint new mon-pipeline # scaffold a valid .blueprint.json
grimoire blueprint validate mon-pipeline.blueprint.json
grimoire blueprint compile mon-pipeline.blueprint.json
grimoire blueprint evals mon-pipeline.blueprint.json --record trace.json
Evals attached to a blueprint describe what the flow must produce, not just the
shape it must have. The Studio never runs them: your host runs the flow once and
records what happened (contract returned, tokens spent, verdict, path taken),
then blueprint evals checks the recording. A case missing from the record is
reported as "not executed" — never as a failure.
- Reference schema:
schemas/blueprint-v1.schema.json - Ready-to-use examples:
registry/blueprints/(minimal,web-pipeline) - The visual studio remains available through
grimoire serve(the multi-project cockpit is served bygrimoire cockpit)
Every validation error reports the offending JSON path, the expected value and
the remediation; a missing extension at compile time prints the exact
grimoire ext add command.
Use the Python SDK¶
from grimoire.core.config import GrimoireConfig
from grimoire.core.project import GrimoireProject
config = GrimoireConfig.from_yaml("project-context.yaml")
project = GrimoireProject(config)
# List the agents
for agent in project.status().agents:
print(f"{agent.id}: {agent.name}")
MCP server¶
If you installed grimoire-kit[mcp]:
Configure it in VS Code (.vscode/mcp.json):
Shell completion¶
Grimoire Kit supports completion for Bash, Zsh and Fish:
# Install completion for your shell
grimoire --install-completion
# Print the script without installing it
grimoire --show-completion
After installing, restart your terminal. Type grimoire <TAB> to see the
available commands.
Next steps¶
- YAML reference — full schema of
project-context.yaml - SDK guide — using the Python SDK
- MCP integration — MCP server for Copilot
- Migration v2 to v3 — migrate an existing project
- Concepts — architecture and principles