Skip to main content

Methods

A method is a learned workflow. When an agent finishes a multi-step task, it can save the sequence of tool calls as a YAML body that future tasks reuse. Methods sit alongside Skills and Tools in the same RAG index — the preflight search picks all three together.

This page is the developer reference: what's in a method, how scoring works, where they're stored, and how the auto-selection mechanics around methods/tools/skills are tuned.

What's in a method

PropertyDescription
nameDisplay name
descriptionBrief description (used by RAG search)
triggerTextOptional phrases that activate this method ("deploy to staging", "publish post")
bodyThe YAML workflow — step-by-step tool calls with logic between them
toolsUsed / skillsUsedAuto-extracted from body so loading the method auto-loads its dependencies
tokenCountEstimated cost for context budgeting
versionBumped on every edit

Loading a method automatically loads its referenced tools and skills. So an agent that starts with research skills can pull in a "deploy to staging" method (and the tools that method uses) without you doing anything.

Scoring

Methods are scored using a recency-weighted success rate:

score = successRate × recencyWeight
recencyWeight = 1.0 / (1.0 + daysSinceUsed / 30.0)

Each time a method is used, the system records a MethodEvent (loaded, succeeded, failed) and recalculates the score. High-quality, recently-used methods rank higher in search results — so the workflows that actually work float to the top automatically.

Before every chat message, a preflight RAG search runs across all enabled skills, methods, and tools. It uses hybrid BM25 + vector matching to find items relevant to the user's query, then injects matching skill instructions and method bodies into the system prompt.

The search runs through the configured Core Model (Settings → General → Core Model), and falls back to the active chat model when Core Model is unset. A small fast Core Model (foundation on macOS 26+, or gemma-4-e2b-it-4bit) keeps preflight cheap.

Search width controls how aggressively it pulls candidates per index:

ModeMethodsToolsSkillsBest for
off000Disabling auto-selection entirely
narrow121Fastest responses, minimal context
balanced (default)352Most cases — good coverage at moderate cost
wide584Maximum coverage, larger prompts

Set the mode in Management → Settings → Capabilities.

Mid-conversation discovery

The agent can also expand its kit while a chat is in progress via two always-on tools:

ToolWhat it does
capabilities_searchSearch methods, tools, and skills across all indexes in parallel
capabilities_loadLoad a specific capability by ID into the active session

Loading a method through capabilities_load automatically loads its referenced tools and skills.

Storage

Methods live in ~/.osaurus/methods/methods.sqlite (encrypted with SQLCipher since 0.17.7).

Browsing methods

There isn't a dedicated "Methods" tab in the app — methods live alongside skills in the same RAG index. To inspect what's been learned:

  • Management → Insights shows when methods were loaded and whether they succeeded or failed
  • The methods database is browsable via SQLite tools if you really need to dig in

Skills, Methods, Tools — the comparison

SkillsMethodsTools
SourceYou author them or import from a marketplaceAgents save them after successful runsBuilt-in plugins, native plugins, MCP providers
ContentMarkdown instructions + reference filesYAML sequences of tool callsCode (Swift, Rust, Python via MCP)
What they doAdd domain knowledge / methodologyReplay a known-good workflowTake action (read files, run commands, call APIs)
Loaded byRAG search (preflight + on-demand)RAG search (preflight + on-demand)RAG search; loading a method auto-loads its tools
Token costThe skill's instructions textThe method's YAML bodyJust the tool's spec (description + parameters)

Related:

  • Skills — user-facing skills documentation
  • Tools & Plugins — what tools exist and how they're built
  • Agents — capabilities are scoped per agent