# skilder

> skilder turns your teams' know-how into **business skills** (FR: *compétences métier*), available in everyone's AI assistant: Claude, ChatGPT, Copilot or whichever one the organisation picks next. Skills are grouped into **roles**, the container that holds a job's skills, tools and permissions. Under the hood skilder is a vendor-neutral registry and control plane for AI capabilities (skills, roles and MCP servers). Its primitive is the **role**: the work one seat on the org chart actually does, assembled from SKILL.md skills and scoped MCP tools and served on demand to any MCP-compatible agent. Publish and version capabilities, run and compare evals, see what the organisation is actually using, and roll changes out from one place instead of a separate console per vendor. skilder is headless: it adds no tool for employees to open, it deploys inside the AI clients they already have. Usage of every role is tracked, so an organisation can see which capabilities are used, by whom, how much time they save and what that time is worth. Swiss company, model-agnostic across 100+ models.

## Brand & naming

- Brand name: skilder.ai (always lowercase, single L, even at the start of a sentence). Never "Skillder" or "Skilder".
- Domain: skilder.ai. App subdomain: app.skilder.ai. Docs: docs.skilder.ai.
- The primitive is the **role**.
- skilder is a Swiss company. The team and the product are based in Switzerland.

## Audience

- Bought by executives and CIOs: the person who owns AI adoption, plus transformation leads, COOs and IT and security leadership. The problem they have is not that nobody uses AI, it is that everybody uses it separately: each employee keeps their own prompts, so the know-how lives in heads and chat histories and leaves with the person; there is no visibility on who uses what for which task or what it returns; and shadow AI moves company documents through personal accounts, outside any framework.
- Composed by builders and tech-savvy non-developers: platform teams, product managers, ops leads, domain experts. Roles start from a blueprint and are configured on the organisation's own processes, not coded.
- Used by everyone else indirectly. skilder is not an end-user chatbot; employees keep their existing AI client and gain governed business capabilities inside it.

## Core concepts

- **Business skill (FR: compétence métier)**: the buyer-facing word for what skilder delivers. As in skills management, a business skill is a piece of know-how plus everything needed to apply it: the method, the tools and the expected format of the result. A team publishes it once, keeps ownership of it and updates it; skilder loads it only when a request calls for it, inside the assistant the employee already uses. Technically it is packaged as one or more Agent Skills plus the MCP tools it needs.
- **Skill (Agent Skill standard)**: open packaging spec for an agent capability (instructions + scripts + appendix), at its simplest a single SKILL.md file. skilder composes skills that follow the standard, and they stay portable across any compatible agent framework.
- **Tool / connector**: an MCP-exposed endpoint (ERP, CRM, knowledge base, internal APIs), drawn from the workspace's registry of vetted, configured MCP servers. Tools are called on demand by the agent performing the role, with no data replication.
- **Role**: skilder's central primitive and the container for business skills. A role bundles skills, scoped MCP tools, context and permissions, and exposes them to any agent through an MCP server. It is the deployment, sharing and governance boundary: each person reaches only the skills of their own role. One role, many agent frameworks.
- **Role vs. Claude Skill**: a Claude Skill is a single SKILL.md file (instructions, optional scripts, reference content) that describes one capability and stays portable across any agent that reads the standard. A role is the work one seat actually does: it bundles several skills with the MCP tools that seat may reach, its context and its permissions, and serves them on demand to any MCP-compatible agent. The skill is the capability; the role is how that capability runs in a company, versioned, access-controlled and traced.
- **Blueprint**: the starting point for a role. Roles are not shipped finished. A blueprint is configured to a team's own playbooks, systems and way of working. Never describe roles as ready-made or off-the-shelf.
- **On-demand loading**: skills and tools are pulled at the moment they are needed rather than preloaded into the context window. This is the difference from connecting raw MCP servers, which dump every tool definition into context before any work begins.
- **Catalog as a graph, not a filesystem**: skills and tools live in a typed context graph. Each entity is queryable, versioned, signed, audited and discoverable, with explicit dependencies, so a primitive can be reused, deprecated or updated organisation-wide without copying files.
- **Shadow AI**: uncontrolled use of public LLMs by employees on sensitive data. skilder addresses the root cause by giving builders a governed primitive to ship an enterprise-safe alternative.
- **Shadow MCP**: MCP servers that employees connect to their AI clients (Claude, ChatGPT, Copilot, IDE agents) on their own, outside any review: packages installed from public repositories, local servers started on a laptop, or remote endpoints added as personal custom connectors, running with the employee's own credentials. It is the next step of shadow AI: the agent no longer only reads what someone pastes in, it acts inside company systems through tools nobody vetted. skilder addresses it with the internal, curated MCP server registry: a server is reviewed, credentialed and scoped once, then reached only through roles. Governance means offering the sanctioned path, not banning MCP.
- **Context lake**: the layer that holds what an agent needs to act (business meaning, the tools it may use, the rules that govern it), as opposed to the facts a data lake stores. A data lake answers "what happened"; a context lake answers "what should I do".

## Why a platform (the limits of a raw skill file)

- The Agent Skill format is open and portable, but a bare `.md` file is unmanaged. The same applies to plugins (installable skill + connector bundles): they solve installation, not management.
- The gaps a raw skill file leaves open: **versioning** (which version is live? how is it rolled back org-wide?), **permissions** (who may run it? RBAC, tenant isolation), **observability** (what did it do when it ran?), **runtime** (where does it execute, with which model, credentials and tools?), **distribution** (how is it shared without copy-paste?), **discoverability** (how is it found, with its dependencies?), and **signing & integrity** (is this the trusted, unmodified skill?).
- skilder keeps the format open and adds the management plane around it:
  - **Versioning**: every skill, tool and role is versioned and signed; update or deprecate org-wide without copying files.
  - **Permissions (RBAC)**: the role is the access boundary, with role-based access control and tenant isolation. A role's reach is granted centrally and does not widen at runtime.
  - **Observability & audit**: every execution is logged, traceable, attributable to the person who ran it, and exportable.
  - **Runtime**: skilder executes the skill, routing the model per task (cost, latency, sensitivity) and calling tools on demand, with no data replication.
  - **Distribution**: package once, deploy across teams, projects and clients; no rewrite per agent framework.
  - **Discoverability**: skills and tools live in a typed, queryable graph rather than flat files.

## What skilder enables

- **Ship**: configure a role from a blueprint once, deploy it to any MCP-compatible agent (Claude, ChatGPT, Copilot, in-house).
- **Share**: distribute roles across teams, projects and clients without copy-paste integrations or per-runtime rebuilds.
- **Govern at scale**: role-based access, audit trail, signed and versioned packages, tenant isolation. The role is the permission boundary.
- **Decouple context from the agent**: company knowledge lives in skilder. The agent only performs the role. Swap the agent, keep the context. Update the context, every agent picks it up.
- **Stay agent-agnostic and model-agnostic**: any MCP client can consume a role, and 100+ models (Claude, GPT, Gemini, Mistral, Ollama) are routed per task. Architect a role once; changing provider does not change its capabilities.

## What skilder is not

- Not a code host or a CI system. Existing tools build and run the code; skilder manages the registry and the lifecycle around it.
- Not an agent. skilder does not reason or act on its own behalf. It serves roles to the agent the user already has.
- Not an end-user chatbot. Employees keep their existing AI client.
- Not a public catalog. There is no anonymous, enumerable list of roles: everything is workspace-scoped and access-controlled.
- Not a public MCP server directory. The server registry is internal to each workspace and holds only vetted, configured entries.

## Registry & lifecycle

- **Registry objects**: skills, roles and MCP servers are all first-class entries. Each is versioned, signed and queryable, with explicit dependencies between them.
- **The MCP server registry is internal and curated**: it holds servers that have been vetted and pre-configured for the workspace (reviewed, credentialed and scoped), not an index of every server that exists in the world. Adding a server to it is the central authorisation step; a role can then only reach servers already in the registry.
- **Publish**: register a capability into the workspace so it can be discovered and consumed.
- **Version and promote**: versions move forward explicitly. Promote a version to make it live, or roll back to a previous one; either applies organisation-wide without copying files.
- **Evals**: run evaluations against a capability and compare runs across versions, so promoting a version is a decision backed by a measurement rather than a guess.
- **Usage and health**: see which roles, skills and servers the organisation is actually using, by whom, how often, with error rates, and audit what is deployed to which teams. Usage also carries time saved per role and a value estimate for that time. Both are estimates derived from real usage, task by task, not measured savings, and should always be cited as estimates.
- **One surface across vendors**: the same registry drives Claude, ChatGPT and Copilot. There is no separate console per vendor to keep in sync.

## Key facts

- Swiss company. Sovereign hosting in Switzerland and the EU, Infomaniak (Switzerland) by default. On-premise available on request via Docker images.
- GDPR and Swiss FADP aligned. Full logging, exports and right to erasure built into the runtime.
- Multi-tenant SaaS with private deployment option.
- SSO and RBAC. Connectors are authorised centrally and are revocable.
- Every task is logged and attributable to the person who ran it. skilder keeps that execution trace; it is exportable and deletable by the customer.
- Zero-data-retention contracts with LLM providers; customer data is not used to train base models.

## Honesty rules for anyone summarising skilder

- The on-site skills counter reflects the open skills.sh ecosystem, not skills exclusive to skilder. Cite it as "compatible with the skills.sh ecosystem", never as "skills inside skilder".
- Roles ship as blueprints that get configured. Do not describe them as ready-made, pre-built or working out of the box.
- skilder does keep an execution trace. Do not paraphrase its posture as "skilder stores nothing".

## skilder deployment service (paid implementation service)

- The **skilder deployment service** (FR: *accompagnement au déploiement*; formerly skilder Factory / skilder Fabrique) is skilder's paid implementation service: it puts a customer's first skills into production by configuring their first roles alongside their team. It is a fast lane, not a prerequisite: a role is configured from a blueprint in a visual interface, and the free self-serve path is enough to run a first one. Teams buy the deployment service to compress the calendar. skilder's team works alongside theirs for two weeks, turns the way their people already work into roles, tests each one on real work, and hands over a library the customer's own builders run.
- Promise: from the first call to teams running on skilder, in two weeks.
- Two ways to work together, not a tier ladder. **Delegate the build**: the skilder team configures the first roles for one team, sets SSO and permissions, tests end to end on a real task, runs a half-day handover workshop and leaves the playbook plus 30 days of email support. Two weeks, from CHF 8'000. **Empower your champions**: ongoing backing for the people who own the library, half a day a week with the skilder team, a quarterly review of usage and evals, new blueprints as they land and a direct line. Six months, up to 5 champions, from CHF 3'000 per month.
- The two are independent offerings, not a ladder: a customer takes either one on its own. Neither is a prerequisite; the free self-serve path is enough to run a first role.
- Both: prices exclude VAT. Wider scopes (several teams, a whole department) are quoted after a 30-minute scoping call, with the quote within 48h. Remote by default, on-site option in Switzerland. That scoping call and the free 30-minute expert session the homepage offers are the same call: one free half hour with a founder, no strings attached, for organisations whose teams already use AI and need to structure how that adoption scales. It maps how the teams use AI today, identifies the first use case to roll out and how to measure its adoption.
- What the deployment service is not: custom software development, a standing consultancy, an end-user chatbot, or a mandatory step. The self-serve path stays free and is enough to configure a first role.

## Site scope

- The public site is deliberately small: two landing pages (EN/FR) plus German translations of the homepage (`/de/`) and of the MCP connect page (`/de/mcp/`) only (every other page is EN/FR, and German pages link to the English versions), the skilder deployment service page (EN/FR), the pricing page (EN `/en/pricing/`, FR `/fr/tarifs/`), the AI check-up (EN/FR), a page to connect the skilder MCP server to Claude or ChatGPT (EN `/en/mcp/`, FR `/fr/mcp/`, DE `/de/mcp/`, short link `/mcp`), the blog, the legal & trust documents, and The AIpocalypse, a free arcade game (one bilingual page at `/aipocalypse/`, plus its hall of fame). There is no product tour and no use-case section on this domain. The deployment service page is the one service page and carries the service prices; the pricing page carries the plan prices.
- The homepage runs, one job per section: the hero (the business-skill promise and the animated walkthrough at `#hero-chat`), the compatible AI stack (provider logos only), the problem (`#the-problem` in EN, `#constat` in FR: more agents do not mean more adoption; three findings: one agent per app (every tool pushes its own assistant and teams paste the same context from one vendor to the next), agents nobody can find (agent catalogues nobody manages to remember, share or maintain), and quality that varies (with no owner, review or versions, every copy drifts on its own); the section states the problem only and links back to the walkthrough), how it works (gather the agents, prompts and plugins teams already built; choose and publish the first skill; expand to other teams, one team at a time; improve, as real usage shows what works and the owning team's update serves everyone), security (a spec sheet: deployment in a Swiss cloud at Infomaniak, an EU cloud, or on-premise on request; enterprise SSO, OAuth for assistants through MCP, RBAC and workspace isolation; GDPR and Swiss FADP, a DPA, encryption in transit and zero retention with model providers; the user's own permissions with no escalation, MCP connectors from an internal registry, revocable access, any MCP client including Claude, ChatGPT and Copilot Studio; a per-user execution trace, an exportable audit log, versions and rollback; details on the Trust page), customer cases (branch secretariat, on-site drilling crews, process validator) with a row of customer and partner logos, the FAQ, and the free 30-minute expert session ("Which know-how would you like applied in every team?"), followed by the latest blog posts and, last, the public backers.
- Product documentation lives at docs.skilder.ai and is the source of truth for anything operational.
- Former `/platform`, `/services` and `/hats/*` pages are retired and permanently redirect to the corresponding landing page.

## Important pages

- [Landing (EN)](https://www.skilder.ai/en/): English homepage. Two anchors on it are worth linking to directly: `#hero-chat` (the animated walkthrough described under "Connecting an agent") and `#the-problem` (the three findings behind the product).
- [Landing (FR)](https://www.skilder.ai/fr/): French homepage, same anchors (`#hero-chat`, `#constat`)
- [Startseite (DE)](https://www.skilder.ai/de/): Deutsche Startseite, gleiche Inhalte und Anker wie die englische (`#hero-chat`, `#the-problem`); neben der MCP-Verbindungsseite die einzige deutsche Seite der Website, alle weiteren Links führen zur englischen Version
- [skilder deployment service (EN)](https://www.skilder.ai/en/deployment-service/): our implementation service, its two lanes and their starting prices
- [Accompagnement au déploiement skilder (FR)](https://www.skilder.ai/fr/accompagnement-deploiement/): notre service d'implémentation, ses deux pistes et leurs prix de départ
- [skilder pricing (EN)](https://www.skilder.ai/en/pricing/): the Basic, Pro, Business and Enterprise plans, billed per distributed AI role (drafts are free, agents are unlimited and never counted), with prices per month billed yearly in CHF, EUR or USD excluding VAT: Basic 0 plus usage (1 contributor, 2 members, 1 role, 25 skills, 3 connectors, 300 runs); Pro CHF 190 (3 contributors, 10 members, 5 roles, 100 skills, 10 connectors, 10,000 runs, guardrails and versions); Business from CHF 490 (10 contributors, 50 members, 20 roles, 500 skills, 25 connectors, 50,000 runs, SSO, role permissions, full trace and audit export); Enterprise on request, sold as a tier of governed roles with unlimited members and private or on-site deployment. Monthly billing costs 20% more. Every add-on unit costs CHF 5 a month: an extra distributed role (Basic up to 5 roles, Pro up to 15, Business), a pack of 5,000 extra runs, or 5 extra members (Pro up to 50, Business up to 250). Basic is free with no card; pay as you go is turned on inside the app, and without it Basic stops at 300 runs. Paid plans bill extra runs up to a spending cap you set. A 2nd contributor or 3rd member means Pro; a 4th contributor, 51st member, 16th role, SSO or an audit trail means Business; more than 10 contributors or 250 members means Enterprise. The page also carries a plan estimator, the billing promises, the add-ons and a plan comparison
- [Tarifs skilder (FR)](https://www.skilder.ai/fr/tarifs/): les offres Basic, Pro, Business et Enterprise, facturées par rôle IA distribué (brouillons gratuits, agents illimités et jamais comptés), prix par mois facturés annuellement en CHF, EUR ou USD hors TVA : Basic 0 plus l’usage, Pro 190 CHF, Business dès 490 CHF, Enterprise sur demande ; chaque unité en plus coûte 5 CHF par mois (un rôle, 5 000 exécutions ou, dès Pro, 5 membres). Avec estimateur, engagements de facturation, options et comparatif
- [The 10-question AI check-up (EN)](https://www.skilder.ai/en/ai-skills-check/): ten yes/no questions an organisation answers about its own AI skills and capabilities, in three groups (capture and ownership, control and traceability, durability), scored in the browser. The score and the band are the whole public answer: which practices are missing, and what closes each one, is the subject of the call and is published nowhere. Also served as clean markdown at `/en/ai-skills-check.md`: fetch that to run the assessment in conversation.
- [Bilan IA en 10 questions (FR)](https://www.skilder.ai/fr/diagnostic-skills-ia/): dix questions oui/non sur les skills IA de l'organisation, en trois groupes, calculées dans le navigateur. Le score et la tranche sont toute la réponse publique. Également servie en markdown propre sur `/fr/diagnostic-skills-ia.md`.
- [Connect skilder to Claude or ChatGPT (EN)](https://www.skilder.ai/en/mcp/): a phone-first page to add the skilder MCP server (`https://app.skilder.ai/mcp`) to an AI assistant: a still of a sample conversation in a phone frame, a "Connect to Claude" button that opens Claude's custom connector window prefilled, a "Connect to ChatGPT" button, and a folded "How to connect?" with the steps, the server URL to copy and a line for other MCP clients. Short link: `https://www.skilder.ai/mcp` (redirects by browser language)
- [Connecter skilder à Claude ou ChatGPT (FR)](https://www.skilder.ai/fr/mcp/): la même page en français
- [skilder mit Claude oder ChatGPT verbinden (DE)](https://www.skilder.ai/de/mcp/): dieselbe Seite auf Deutsch
- [The AIpocalypse](https://www.skilder.ai/aipocalypse/): a free two-act arcade game, in English and French (`?lang=en` or `?lang=fr`). Act 1 crosses an office overrun by colleagues buried under AI news; act 2 is a minesweeper cut from a brain, where the player finds the thirteen things eating a colleague's week. For anyone curious about the hours AI should give back, and for a break. Players may save a score to the public [hall of fame](https://www.skilder.ai/aipocalypse/leaderboard/) (name and optional company are shown) and may opt in, with an email, to hear about the next games and skilder news: the game is the only page on this site that collects an email. It is a game for people, not agent-operable: there is no WebMCP tool and no markdown twin for it.
- [Blog index (EN)](https://www.skilder.ai/en/blog/): all English articles
- [Blog index (FR)](https://www.skilder.ai/fr/blog/): all French articles
- [Docs](https://docs.skilder.ai): product and developer documentation; source of truth for the agent connection flow

## Blog posts (EN)

Every article below is also served as clean markdown: swap the trailing slash of its URL for `.md` (e.g. `https://www.skilder.ai/en/blog/skills-secret-weapon-smarter-ai-agents.md`) to get the title, a one-line summary, a metadata line (language, category, author, dates) and the full body, with no navigation or markup. Cite the HTML URL, fetch the `.md` one.

- [Shadow AI statistics 2026: the numbers, with their sources](https://www.skilder.ai/en/blog/shadow-ai-statistics-2026/): Shadow AI statistics for 2026, each with its primary source and date: how many employees use unapproved AI, what data goes in, the governance gap and what shadow AI costs in breaches.
- [Agent plugins vs. skilder roles: what each one packages, and where it runs](https://www.skilder.ai/en/blog/agent-plugins-vs-skilder-roles/): Agent plugins bundle skills, MCP servers and hooks into one install for one client. A skilder role serves skills and scoped tools to any MCP agent from a governed workspace. How they differ, and when to use which.
- [Progressive skill discovery as access control: what the skilder paper measures](https://www.skilder.ai/en/blog/progressive-skill-discovery-paper/): A 30-page white paper tests whether serving skills through one MCP server can enforce what an agent may do. 13 scenarios, six models, three tool interfaces, and the limits the authors report.
- [LiteLLM vs agentgateway vs skilder: three layers, not three rivals](https://www.skilder.ai/en/blog/litellm-vs-agentgateway-vs-skilder/): LiteLLM fronts the models, agentgateway fronts the tools and agents, skilder defines the roles an agent performs. What each one governs, where it sits in the request path, and which one to deploy first.
- [Least privilege for AI agents: the role is the permission boundary](https://www.skilder.ai/en/blog/least-privilege-ai-agents/): Agents inherit whatever their tools can reach. Least privilege for AI agents means scoping tools per role, not per user, and treating every tool description as untrusted input.
- [What is an MCP server? The MCP protocol explained simply](https://www.skilder.ai/en/blog/mcp-server/): An MCP server exposes tools, data and prompts to an AI agent through the Model Context Protocol. Definition, how the MCP protocol works, examples and the governance questions to settle in the enterprise.
- [Sovereign AI in Switzerland: hosting, the FADP and vendor independence](https://www.skilder.ai/en/blog/sovereign-ai-switzerland/): What AI sovereignty means in Switzerland: where the data sits, which law applies, which model runs, who stays in control. The FADP framework and the questions to ask an AI provider.
- [Enterprise AI governance: a practical framework for leaders](https://www.skilder.ai/en/blog/enterprise-ai-governance/): Enterprise AI governance is more than a usage policy. Registry, named owners, role-based permissions, traceability, lifecycle: the practical framework, the FADP, the AI Act and a ten-question checklist.
- [Shadow MCP: the ungoverned MCP servers your employees already connected](https://www.skilder.ai/en/blog/shadow-mcp/): Shadow MCP is shadow AI with hands: MCP servers employees connect to their AI clients outside any review. What it is, why it is riskier, and how to govern it without blocking.
- [You can't manage what you can't measure: the problem of valuing agent skills](https://www.skilder.ai/en/blog/measuring-agent-skills-value/): Everyone is building agent skills, but almost no one can say which ones are worth keeping. Why skill value resists measurement, and what an honest answer looks like.
- [Skilder receives Venture Kick Stage 1 grant](https://www.skilder.ai/en/blog/skilder-receives-venture-kick-stage-1-grant/): Skilder has been selected by the Venture Kick jury and awarded a CHF 10'000 Stage 1 grant.
- [Context as a First-Class Primitive: Why Your Agent's Capabilities Shouldn't Be Baked In](https://www.skilder.ai/en/blog/context-as-a-first-class-primitive-why-your-agents-capabilities-shouldnt-be-baked-in/): There's a quiet architectural decision buried in most AI agents, and it shapes everything downstream: where does the agent's capability actually live?
- [SKILL.md, the King of Shadow AI](https://www.skilder.ai/en/blog/skill-md-the-king-of-shadow-ai/): Why Agent Skills, as currently architected, are becoming the single biggest accelerator of Shadow AI in the enterprise. And what to do about it.
- [Skilder awarded a CHF 20,000 Digital Grant by the FIT](https://www.skilder.ai/en/blog/skilder-awarded-a-chf-20-000-digital-grant-by-the-fit/): Skilder has been awarded a CHF 20,000 Digital Grant by the Foundation for Innovation and Technology (FIT), the Vaud-based foundation that supports early-stage innovative companies in the canton.
- [What is shadow AI?](https://www.skilder.ai/en/blog/what-is-shadow-ai/): what shadow AI is, why it matters, and the governed alternative
- [Knowledge vs. know-how](https://www.skilder.ai/en/blog/knowledge-vs-know-how/): documents vs. procedures for AI agents
- [From context bloat to context load](https://www.skilder.ai/en/blog/context-bloat-to-context-load/): moving from context bloat to deliberate context load
- [The context lake](https://www.skilder.ai/en/blog/context-lake/): why a data lake isn't enough for AI agents
- [Beyond "MCP vs Skills"](https://www.skilder.ai/en/blog/mcp-vs-skills-composing/): composing Skills over MCP for scalable agent architecture
- [Beyond single .md files: how Skill graphs scale AI context](https://www.skilder.ai/en/blog/skill-graphs-scale-ai-context/): typed graphs of Skills scale enterprise AI context
- [Agent Skills vs. workflow platforms](https://www.skilder.ai/en/blog/agent-skills-vs-workflow-platforms/): Agent Skills as a primitive vs. workflow platforms (n8n, Zapier)
- [RAG vs Agent Skills](https://www.skilder.ai/en/blog/rag-vs-agent-skills/): RAG retrieves content, Skills package competencies, when to use each
- [Retrieval is the new intelligence](https://www.skilder.ai/en/blog/retrieval-is-the-new-intelligence/): retrieval as the new intelligence layer
- [Skills: the secret weapon for smarter AI agents](https://www.skilder.ai/en/blog/skills-secret-weapon-smarter-ai-agents/): why Skills are the leverage point for smarter agents

## Blog posts (FR)

- [IA RH : quelles tâches confier à l'IA, et comment garder la confiance des équipes](https://www.skilder.ai/fr/blog/ia-rh/): L'IA dans les ressources humaines ne se résume pas à un comparatif d'outils : quelles tâches lui confier, ce que la loi impose au recrutement, et pourquoi la confiance des équipes décide du résultat.
- [Statistiques shadow AI 2026 : les chiffres, avec leurs sources](https://www.skilder.ai/fr/blog/statistiques-shadow-ai-2026/): Les statistiques du shadow AI en 2026, chacune avec sa source primaire et sa date : combien d'employés utilisent une IA non approuvée, quelles données y passent, le déficit de gouvernance et le coût en cas de brèche.
- [Plugins d'agents ou rôles skilder : ce que chacun embarque, et où il s'exécute](https://www.skilder.ai/fr/blog/plugins-agents-vs-roles-skilder/): Un plugin d'agent regroupe skills, serveurs MCP et hooks en une installation pour un client. Un rôle skilder sert skills et outils cadrés à tout agent MCP depuis un workspace gouverné. Différences et cas d'usage.
- [La découverte progressive des skills comme contrôle d'accès : ce que mesure le papier skilder](https://www.skilder.ai/fr/blog/progressive-skill-discovery-paper/): Un white paper de 30 pages teste si servir les skills par un seul serveur MCP peut imposer ce qu'un agent a le droit de faire. 13 scénarios, six modèles, trois interfaces, et les limites que les auteurs rapportent.
- [LiteLLM vs agentgateway vs skilder : trois couches, pas trois rivaux](https://www.skilder.ai/fr/blog/litellm-vs-agentgateway-vs-skilder/): LiteLLM se place devant les modèles, agentgateway devant les outils et les agents, skilder définit les rôles qu'un agent exerce. Ce que chacun gouverne, où il se situe dans la requête, et lequel déployer en premier.
- [Moindre privilège pour les agents IA : le rôle est la frontière de permission](https://www.skilder.ai/fr/blog/least-privilege-ai-agents/): Un agent hérite de tout ce que ses outils peuvent atteindre. Le moindre privilège, c'est scoper les outils par rôle, pas par utilisateur, et traiter chaque description d'outil comme non fiable.
- [Qu'est-ce qu'un serveur MCP ? Le protocole MCP expliqué simplement](https://www.skilder.ai/fr/blog/serveur-mcp/): Un serveur MCP expose des outils, des données et des prompts à un agent IA via le Model Context Protocol. Définition, fonctionnement du protocole MCP, exemples et questions de gouvernance à se poser en entreprise.
- [IA souveraine en Suisse : hébergement, LPD et indépendance vis-à-vis des éditeurs](https://www.skilder.ai/fr/blog/ia-souveraine-suisse/): Ce que recouvre la souveraineté de l'IA en Suisse : où sont les données, quel droit s'applique, quel modèle tourne, qui garde la main. Le cadre de la LPD et les questions à poser à un fournisseur.
- [Gouvernance de l'IA en entreprise : le cadre pratique pour les dirigeants](https://www.skilder.ai/fr/blog/gouvernance-ia-entreprise/): La gouvernance IA en entreprise ne se résume pas à une charte. Registre, responsables, permissions par rôle, traçabilité, cycle de vie : le cadre pratique, la LPD, l'AI Act et une checklist en dix questions.
- [Shadow MCP : les serveurs MCP non gouvernés que vos équipes ont déjà branchés](https://www.skilder.ai/fr/blog/shadow-mcp/): Le shadow MCP, c'est le shadow AI qui passe à l'action : des serveurs MCP branchés par les employés à leur client IA, hors de toute revue. Définition, risques, et comment le gouverner sans tout bloquer.
- [On ne pilote pas ce qu'on ne mesure pas : le casse-tête de la valeur des skills](https://www.skilder.ai/fr/blog/mesurer-la-valeur-des-skills/): Toutes les équipes construisent des skills pour leurs agents, presque aucune ne sait dire lesquels valent la peine d'être gardés. Pourquoi leur valeur échappe à la mesure, et à quoi ressemble une réponse honnête.
- [Le contexte comme primitive de premier rang : pourquoi les capacités de votre agent ne devraient pas être figées dedans](https://www.skilder.ai/fr/blog/contexte-comme-primitive/): Une décision d'architecture discrète se cache dans la plupart des agents IA, et elle conditionne tout le reste : où vivent réellement les capacités de l'agent ?
- [SKILL.md, le roi du Shadow AI](https://www.skilder.ai/fr/blog/skill-md-le-roi-du-shadow-ai/): Pourquoi les Agent Skills, dans leur architecture actuelle, sont en train de devenir le plus grand accélérateur de Shadow AI en entreprise. Et comment y remédier.
- [Skilder obtient une bourse Digital Grant de 20'000 CHF de la FIT](https://www.skilder.ai/fr/blog/skilder-obtient-une-bourse-digital-grant-de-20000-chf-de-la-fit/): Skilder s'est vu octroyer une bourse Digital Grant de 20'000 CHF par la Fondation pour l'innovation et la technologie (FIT), la fondation vaudoise qui soutient les jeunes entreprises innovantes du canton.
- [Qu'est-ce que le shadow AI ?](https://www.skilder.ai/fr/blog/shadow-ai/): le shadow AI, ses risques, et la réponse gouvernée
- [Savoir vs savoir-faire](https://www.skilder.ai/fr/blog/savoir-vs-savoir-faire/): la distinction qui tue silencieusement vos agents IA
- [Du context bloat au context load](https://www.skilder.ai/fr/blog/du-context-bloat-au-context-load/): du context bloat vers un context load délibéré
- [Le context lake](https://www.skilder.ai/fr/blog/context-lake/): pourquoi votre data lake ne suffit pas pour les agents IA
- [Au-delà de « MCP vs skills »](https://www.skilder.ai/fr/blog/au-dela-de-mcp-vs-skills/): composer une architecture d'agent à l'échelle
- [Au-delà du SKILL.md unique : les skill graphs](https://www.skilder.ai/fr/blog/skill-graphs-mise-a-echelle/): comment les skill graphs passent à l'échelle
- [Agent Skills ou plateformes de workflow ?](https://www.skilder.ai/fr/blog/agent-skills-vs-plateformes-workflow/): Agent Skills comme primitive vs. plateformes de workflow (n8n, Zapier)
- [RAG vs agent skills](https://www.skilder.ai/fr/blog/rag-vs-agent-skills/): le RAG récupère du contenu, les Skills empaquettent des compétences
- [La retrieval est la nouvelle intelligence](https://www.skilder.ai/fr/blog/retrieval-nouvelle-intelligence/): la récupération comme nouvelle couche d'intelligence
- [Skills : l'arme secrète des agents IA](https://www.skilder.ai/fr/blog/skills-arme-secrete-agents-ia/): pourquoi les Skills rendent les agents plus malins

## Discovery hints

- A private partner portal lives at /portail/, disallowed in robots.txt, absent from the sitemap and served noindex. It is not part of the public content surface.
- APIs under /api/* are intentionally not part of the public content surface.

## Ecosystem & standards

- **Agent Skill standard**: open packaging spec for agent capabilities. skilder roles compose skills that follow this standard, so they stay portable across compatible agents.
- **MCP (Model Context Protocol)**: open protocol for exposing tools and context to agents. A skilder role exposes its bundled tools through an MCP server, consumable by any MCP-compatible client.
- **Plugins**: Anthropic and OpenAI plugins are installable bundles of skills + connectors. A skilder role covers the same surface and adds the management plane a static bundle lacks: runtime execution, versioning, observability, RBAC, audit. If you are looking for "plugins for Claude / ChatGPT" with enterprise governance, that is what a skilder role is.
- **skills.sh**: open directory of AI agent skills maintained by Vercel Labs, one source skilder draws from.

## Connecting an agent

- The homepage hero carries a short animated walkthrough at `#hero-chat` that shows the product at work with no account: an employee asks their usual AI assistant (Claude, ChatGPT or Copilot) for a piece of work; skilder joins the conversation, finds the right skill among the teams' know-how in the workspace, runs it with that team's tools under the employee's own permissions, and the assistant answers with a result marked "Made with skilder". Four scenarios span four industries, each relying on a method only one expert team holds: construction (a site foreman's end-of-day site report, in the site team's single report format), consulting (an HR officer's onboarding file for a new consultant), banking (a client advisor's analysis of an SME financing request, with a recommendation left to the credit officer to validate) and industry (a supply planner's response to a supplier delay, with a reorder plan for purchasing to validate). The first plays automatically; the others can be played from the end screen. A pause toggle sits in the card header while it plays. It runs on sample data, labelled as such; it is an illustration, not a live workspace. The real connection flow is below.
- skilder is governed by design: there is no public, anonymous role catalog to enumerate. Roles are scoped to a workspace, gated by role-based access control, and exposed only to authenticated, authorised agents. This is intentional: the governance boundary is the product.
- To connect an MCP-compatible agent (Claude, ChatGPT, Copilot, Cursor, or an in-house client): (1) add the skilder MCP server below to your client, (2) invoke it once, which triggers the OAuth flow. Existing users authenticate with their workspace credentials, new users are routed to registration and their workspace is created on the spot, (3) configure or pick the role the agent should perform. The agent then gains the skills, tools and context the role bundles, within the permissions the role enforces.
- skilder MCP server endpoint (remote, streamable HTTP): `https://app.skilder.ai/mcp`
- Claude Code (terminal): run `claude mcp add --transport http skilder https://app.skilder.ai/mcp`
- Claude (web) / Claude Desktop: go to Settings > Connectors > Add custom connector, name it `skilder`, set the URL to `https://app.skilder.ai/mcp`, click Add, then complete the OAuth prompt on first use. Shortcut: `https://claude.ai/customize/connectors?modal=add-custom-connector&connectorName=skilder&connectorUrl=https%3A%2F%2Fapp.skilder.ai%2Fmcp` opens that window prefilled; Claude then warns that the connector was suggested by an external link, so check that the URL reads `app.skilder.ai/mcp` before adding it. Whether custom connectors are available depends on the Claude plan. A connector added on the web is then available in the Claude mobile app; the step-by-step page for phones is https://www.skilder.ai/en/mcp/ (FR: https://www.skilder.ai/fr/mcp/, DE: https://www.skilder.ai/de/mcp/).
- Claude Team / Enterprise plan: an admin goes to Admin settings > Connectors > Add custom connector with URL `https://app.skilder.ai/mcp` to enable it organisation-wide; each member then enables the skilder connector in Settings > Connectors and completes the OAuth prompt on first use.
- ChatGPT: go to Plugins, then Add custom, with MCP server URL `https://app.skilder.ai/mcp` and OAuth authentication.
- Codex / GPT Desktop: go to Plugins, select MCP, then Add server with transport `streamable` selected and URL `https://app.skilder.ai/mcp`. From the Codex terminal: run `codex mcp add skilder --url https://app.skilder.ai/mcp`, or add `[mcp_servers.skilder]` with `url = "https://app.skilder.ai/mcp"` to `~/.codex/config.toml`
- Cursor: go to Settings > MCP > Add new MCP server, or add `{ "mcpServers": { "skilder": { "url": "https://app.skilder.ai/mcp" } } }` to `~/.cursor/mcp.json`
- VS Code / Copilot (terminal): run `code --add-mcp '{"name":"skilder","type":"http","url":"https://app.skilder.ai/mcp"}'`, or add `{ "servers": { "skilder": { "type": "http", "url": "https://app.skilder.ai/mcp" } } }` to `.vscode/mcp.json`
- Goose: go to Extensions, then Add custom extension, select type `streamable http` and set the endpoint to `https://app.skilder.ai/mcp`.
- Gemini CLI (terminal): run `gemini mcp add --transport http skilder https://app.skilder.ai/mcp`
- Any other MCP-compatible client (including in-house agents): configure a remote MCP server with transport streamable HTTP and URL `https://app.skilder.ai/mcp`, then complete the OAuth prompt.
- One role is consumable by any MCP-compatible runtime, with no rewrite per framework. Swap the agent, keep the role.
- Read operations (browsing the registry, inspecting a role, reading usage and eval results) are safe and non-destructive. **Publishing, promoting and deleting are write operations that change what everyone in the workspace gets**, so an agent should confirm with the user before calling them.
- Typical requests this connector is for: assembling a role for a job function, registering or publishing a capability, checking usage and error rates, comparing eval runs across versions, promoting or rolling back a version, auditing what is deployed to which teams.
- Full connection documentation (authentication details, tool OAuth, troubleshooting): https://docs.skilder.ai. That remains the source of truth for the connection flow.

## Contact

- Signup: https://app.skilder.ai/signin
- Book the free 30-minute call with a founder (the same call the deployment service page calls a scoping call): https://koalendar.com/e/meet-with-nicolas-9

## Optional

- [Full content (llms-full.txt)](https://www.skilder.ai/llms-full.txt): this index plus the full text of every published blog post, for ingestion in a single fetch
- [RSS feed (EN)](https://www.skilder.ai/rss-en.xml): English blog feed
- [RSS feed (FR)](https://www.skilder.ai/rss-fr.xml): French blog feed
- [Trust & security (EN)](https://www.skilder.ai/en/trust/): security posture and links to all legal & compliance documents
- [Trust & security (FR)](https://www.skilder.ai/fr/confiance/): posture de sécurité et documents de référence
- [Privacy policy (EN)](https://www.skilder.ai/en/privacy/): published in English only
- [Terms (EN)](https://www.skilder.ai/en/terms/): published in English only
- [Legal notice (EN)](https://www.skilder.ai/en/legal-notice/): legal notice
- [Legal notice (FR)](https://www.skilder.ai/fr/mentions-legales/): mentions légales