Locara

Claude Agent SDK + Agent Skills

What it is: Anthropic’s agent-building stack. The Claude Agent SDK wraps Claude with built-in tool execution, file/shell access, and an agent loop. Agent Skills are folders of instructions, scripts, and resources that agents discover and load dynamically — published as an open standard so that other vendors can adopt the same format. Status: Agent Skills launched October 16, 2025; the open standard for cross-platform Skills was published December 18, 2025. Supported across Claude.ai, Claude Code, Claude Agent SDK, and the Claude Developer Platform. MCP servers integrate alongside Skills. Most relevant to Locara: Skills + MCP together define the dominant 2025–26 contract for “how AI apps extend their capabilities.” Locara’s app/tool/skill abstractions need to take a position on both — being compatible, being parallel, or being incompatible by design.

Background

By mid-2025, Anthropic had three layers in front of users:

  • Claude.ai (the chat product),
  • Claude Code (the agentic CLI / IDE plugin), and
  • Claude Developer Platform (API + Managed Agents).

The connective tissue was MCP (introduced November 2024) for tools/data integration, and system prompts for behavior shaping. The friction was that “teach Claude how to do a complex multi-step workflow” required either inflating the system prompt or building a bespoke tool.

Agent Skills (launched October 16, 2025 via the engineering blog) is Anthropic’s answer: a Skill is a folder containing a SKILL.md file (the instructions), optional scripts the model can execute, and optional resources (templates, schemas, examples). Claude discovers skills, loads them lazily when relevant to the user’s request, and follows the instructions.

The launch shipped with Anthropic-managed skills for Office formats: PowerPoint (.pptx), Excel (.xlsx), Word (.docx), and PDF — useful as both a real capability and a worked example of the pattern. Pro, Max, Team, and Enterprise users got access immediately.

On December 18, 2025, Anthropic published Agent Skills as an open standard with the explicit goal that other agent runtimes (not just Claude) could load the same skill folders. The format is plain — a SKILL.md with frontmatter, optional scripts, optional bundled files. No proprietary serialization.

The Claude Agent SDK is the harness around all of this: it wraps the Claude API with tool execution (read files, run shell, search), MCP server connections, and skill loading, exposing a single agent abstraction. It’s the same harness that powers Claude Code internally, exposed for anyone to build agents on.

Key design decisions

  • Skills as folders, not config. A skill is a directory you can cd into, read, and edit by hand. Plain text. Version control friendly.
  • Lazy loading. Skills are not always in the prompt. The model is shown skill names and descriptions; full content loads only when the model decides to invoke one.
  • Co-existence with MCP. Skills teach workflows; MCP servers provide tools and resources. A skill can call MCP-provided tools. They’re complementary primitives.
  • Open standard from day one of the public release. The Skills format is documented for non-Anthropic runtimes.
  • Agent SDK ships with built-in tools. Read, Write, Edit, Bash, Grep, Glob, WebFetch, WebSearch, plus an agent loop with planning. Not just an API wrapper.
  • Managed Agents — a separate Anthropic-hosted product where Claude runs in Anthropic’s environment with skills/MCP/tools, exposed via API.
  • Beta header tracking. API access uses dated beta headers (skills-2025-10-02, managed-agents-2026-04-01) — a stability signal that the contract is still evolving.

What worked

  • Skills format is dead simple. A folder + a markdown file. Anyone can author one in minutes; no SDK install, no serialization concern.
  • Lazy loading solved the token-cost problem. MCP’s known weakness is that all tool descriptions sit in every prompt. Skills don’t — only names + descriptions are present until the model invokes one.
  • Cross-product portability. Same skill works in Claude.ai, Claude Code, and via the SDK. The Office-formats skills demoed this clearly.
  • Open-standard publication preempted the “Anthropic lock-in” critique. Within weeks, third-party agent runtimes (Continue, Cline, others) announced compatibility plans.
  • Naming was good. “Skills” is the term most agent communities had been using informally; Anthropic standardized the vocabulary.
  • Strong reference content. Anthropic’s anthropic-skills repo on GitHub demonstrates the pattern at scale (docx, pptx, xlsx, pdf, scheduling, more). Real examples, not just specs.

What failed / criticisms

  • Two overlapping primitives now (MCP + Skills) is conceptual overhead. “When do I write a skill vs. an MCP server?” is a recurring question — the answer (“skills for workflow knowledge, MCP for tool/data access”) is right but takes explanation.
  • No standard sandboxing for skills. Skills can include scripts the agent will execute. No mandated isolation; trust model is “review the skill before installing.”
  • Discovery is unsolved. No canonical Skills marketplace yet; the Anthropic-managed skills are bundled, but third-party skills live in scattered GitHub repos.
  • Permission model is per-skill, coarse-grained. Either Claude has the skill or it doesn’t; there’s no per-script gating within a skill.
  • The cross-vendor adoption story is still aspirational. Open standard published in late 2025; real-world cross-runtime compatibility was incomplete by early 2026.
  • Beta header churn. The dated beta-header pattern means downstream consumers must track Anthropic’s release notes and bump versions; not yet “stable.”
  • Skills bias toward English-language instructions. The format is markdown-friendly but the skill-author UX assumes a model that follows natural language well — which works for Claude, less well for smaller open models.

Specific learnings for Locara

  1. MCP and Skills are both standards Locara must support. They’re not competing — they’re complementary. Locara apps should: (a) consume MCP servers (for tools/data); (b) ship Skills (for workflow knowledge). Picking only one leaves capability on the floor.
  2. The Skill format is exactly the right shape for Locara apps’ “behavior” layer. A folder with a SKILL.md, optional scripts, optional resources is already what a Locara app’s behavior bundle should look like. Adopt the format as-is — there’s no benefit to inventing a parallel “Locara playbook” format.
  3. Locara’s manifest layer is the missing capability declaration for Skills. Skills today have no formal capability declaration; the agent runs the skill’s scripts with whatever privileges the harness has. Locara’s manifest + capability scoping is the natural complement: the manifest declares what the skill’s scripts can touch, the runtime enforces it.
  4. Locara is the local-first sandbox layer for both MCP and Skills. Skills include scripts. MCP servers are subprocesses. Both are unsandboxed by default in Anthropic’s reference implementations. Locara’s wasmtime/WASI substrate is exactly the layer that makes both safe to run on a user’s machine without trust-the-author. This is a real wedge — pitch is “MCP and Skills, but actually safe.”
  5. Don’t build a Locara-specific agent harness from scratch. The Claude Agent SDK is high-quality and OSS-friendly. A Locara app that wants agentic behavior can use the SDK with a local model backend (via Locara’s OpenAI-compatible endpoint or by swapping the LLM client). Compete on the runtime/distribution layer, not on the agent loop.
  6. Open Skills marketplace is an opening for Locara. No canonical Skills registry has won. Locara’s curated catalog could host Skills alongside Apps, with the same review pipeline (signed, capability-declared, manifest-validated). “Curated Skills running in Wasm sandbox” is a defensible niche.
  7. Watch beta-header churn. If Locara depends on Skills/MCP beta endpoints, vendor a wrapper that pins versions — don’t expose Anthropic’s beta-header instability to Locara apps directly.
  8. Don’t rebuild “Office files” — adapt Anthropic’s reference skills. The Anthropic-managed skills for docx/pptx/xlsx/pdf are open-source under permissive licensing. Locara apps that need Office-format generation can use them directly, perhaps with Locara’s runtime providing the script-execution sandbox.
  9. The “Claude Code skill” abstraction maps to Locara’s “tool capability.” The way Claude Code uses skills (lazy load, scoped to a task, optional scripts) is a near-perfect template for how Locara apps should expose extensions to the user.

References