Tools API

List Available Tools

GET /api/tools

Execute Tool

POST /api/tools/:name/execute

{
  "parameters": {
    "url": "https://example.com"
  }
}

Built-in Tools

Tool	Description
`browser`	Web browsing and scraping
`file_system`	File read/write operations
`terminal`	Shell command execution
`code_execute`	Code execution in sandbox
`grep_tool`	Ripgrep-powered regex content search (three-tier: ripgrep → mmap → Python), path-grouped densified output (auto-deduplicates paths when ≥5 matches, saving 18-39% tokens), ReDoS protection, non-code file auto-capping
`glob_tool`	File path pattern search under the workspace
`@codebase` mention	Lightweight workspace file/extension overview injected into the user message (use with grep/glob for exploration)
`web_search`	Multi-engine search with BM25/Reranker filtering — returns cleaned, deduplicated snippets
`render_ui`	Declarative in-chat UI (23 A2UI components). Turn1 ~223 tok; full props in `.agent/docs/A2UI_REFERENCE.md` (auto-seeded when enabled). Fail-closed on unknown types
`update_ui_data`	Incremental `data_update` SSE for live panels (progress, metrics, task lists). Deep-merges `data` model — nested fields update without wiping sibling keys. Mounted with `render_ui` when Interactive UI is enabled
`todo_write`	Structured task planning for multi-step objectives. Hard-enforced MAX_TODOS=20 limit prevents over-planning; single `in_progress` concurrency keeps execution focused. Violations auto-corrected with LLM feedback. Opt-in via Planning toggle or auto-enabled on resume when existing todos detected

Default tool groups vs on-demand capabilities

Scenario	Enabled by default	Opt-in
General chat	`web_search` + `memory`	—
Coding / files	Enable file_ops → `glob_tool` + `grep_tool` + file_* mount immediately (Claude Code–style eager, not discover-deferred)	`code_execute` (bash)
MCP / bound skills	—	`discover_capability_tool` → `skill_select_tool` searches already-bound library (`<BoundSkills>`)
External skill marketplace	—	`skill_discovery_tool` search/install from GitHub/skills.sh; Settings Discover tab also available
Skill hygiene	WebUI Curator sweep	—

MCP Tools

External tools connected via MCP are automatically registered alongside built-in tools. MCP tool names are internally prefixed as mcp__{server}__{tool} for namespace isolation — the API accepts both the prefixed and original names.

Two-Route MCP Routing

When you connect MCP servers with many tools, Myrm automatically selects the optimal routing strategy to minimize token usage while preserving full tool access:

Route	Condition	Behavior	User Experience
Direct	Server tools ≤ threshold AND within aggregate budget	Full tool schemas loaded upfront	Zero-latency tool calls
PTC / Skill	Server tools > threshold OR aggregate budget exceeded	Wrapped as a Skill (`skill_search` → `skill_select` → PTC)	On-demand discovery via skill search

Connecting a 50-tool MCP server does not inflate your Turn1 prompt — large servers are demoted to Skills and discovered on demand. Direct MCP tools show their real names in the UI.

Security rules (ALLOW/ASK/DENY) apply to both direct MCP tools and Skill/PTC invocations. A per-tool DENY rule on mcp__gmail__send_email cannot be bypassed via the Skill path.

POST /api/tools/mcp__github__search_repos/execute

Filtering

Per-agent tool filtering is supported via tool_include / tool_exclude in MCP server config:

{
  "tool_include": ["search_repos", "get_issue"],
  "tool_exclude": ["delete_repo"]
}

See MCP Integration > Tool Filtering for details.

​Tools API

​List Available Tools

​Execute Tool

​Built-in Tools

​Default tool groups vs on-demand capabilities

​MCP Tools

​Two-Route MCP Routing

​Filtering

Tools API

List Available Tools

Execute Tool

Built-in Tools

Default tool groups vs on-demand capabilities

MCP Tools

Two-Route MCP Routing

Filtering