Tools available now

Read this reference

Availability rules: Normal mode exposes the registered set; state-changing tools ask for approval interactively. Plan mode uses a strict allowlist and denies every unknown or action-capable tool. Orchestrate mode adds delegation and native file editing tools: the parent may make basic edits, but delegates significant implementation and testing. Auto mode bypasses review for known-safe observations and guarded native workspace mutations, then model-reviews actions with remaining scope or safety risk. Yolo mode keeps the registered set but bypasses approvals. Optional tools appear only after their prerequisite is available.

Marker	Meaning
Always	Registered in both interactive and headless Orcacode runs.
Interactive	Registered only in the terminal interface.
Conditional	Appears only with a key, configured server, or enabled content.
Gated	Requires approval in normal interactive mode.
Plan-safe	Allowed by the plan-mode allowlist.

Workspace files and search

Tool	Use	Availability	Plan mode
`read_file`	Read a UTF-8 file relative to the workspace.	Always	Plan-safe
`grep`	Find literal text and matching lines below the workspace.	Always	Plan-safe
`glob`	Find files using ``, `?`, and `*` patterns.	Always	Plan-safe
`write_file`	Create or replace a workspace file.	Always · gated	Only Markdown directly in `docs/plan/`
`edit_file`	Replace one exact text fragment in a workspace file, or apply an ordered, transactional `edits` batch of exact replacements or appends across existing files.	Always · gated	Only flat `docs/plan/*.md` files; every path in a batch

All paths are workspace-relative; absolute paths and upward escapes are rejected. In plan mode, every target in a batch must qualify. The orchestrate parent can use write_file and edit_file for basic edits. Existing files need a prior read before write_file replaces them. Use edit_file for an exact replacement that fails safely when its old text is absent.

Execution and compute

Tool	Use	Availability	Plan mode
`shell`	Run one command; return stdout, stderr, and exit status.	Always · gated	Interactive only; every command asks for approval
`process`	Start, inspect, write to, poll, or stop persistent/background processes. Spawn can wait for finite work or wake the interactive agent on exit or one literal output match.	Always · gated	Denied
`pykernel`	Run persistent Python; values and imports survive between calls.	Always · gated	Denied
`bun_repl`	TypeScript/JavaScript REPL implementation with persistent state, imports, and top-level `await`. Known issue: the current bun-repl package refuses piped stdin on recent Bun, so this tool is broken there.	Always · gated · Bun on `PATH` (not sufficient on affected versions)	Denied
`subagent`	Spawn a fresh independent worker for one bounded task.	Always · gated	Denied · allowed in orchestrate mode

Completed foreground subagent calls return token usage, runtime, step and tool-call counts, and bounded model/tool timing samples and cumulative durations. Overlapping calls can make cumulative tool time exceed wall time; detached acknowledgements do not contain this telemetry. process.spawn accepts waitForExit for one-call finite work, notifyOnExit (default true for detached processes), and one-shot literal notifyOnMatch for readiness or important output. Interactive notifications resume the model without a polling tool call and remain hidden from the user transcript. Persistent compute tools intentionally retain state during the session; use them for analysis that benefits from one continuous runtime rather than repeated shell invocations.

Coordination, memory, and recovery

Tool	Use	Availability	Plan mode
`memory_search`	Search global and current-workspace memory, or list recent accessible records.	Always	Plan-safe
`memory_manage`	Explicitly save, update, or forget a workspace or global memory.	Always · gated	Denied
`ask`	Request one to three structured multiple-choice clarifications.	Interactive	Plan-safe
`read_tool_result`	Page through original output captured before truncation.	Interactive	Plan-safe

Memory recall itself is automatic and read-only; memory_search is for explicit inspection or a different query. The dedicated Memory guide covers scope, storage, retrieval, and mutations. read_tool_result remains available in the terminal UI even when output truncation is currently disabled, so previously trimmed output can still be recovered. Headless runs stream their events/results directly and do not register the interactive clarification or result-paging tools.

Web and remote capabilities

Tool	Use	Availability	Plan mode
`web_fetch`	Fetch a public HTTP/S URL and return text or converted HTML.	Always	Plan-safe
`web_search`	Search the web through Firecrawl.	Conditional · `FIRECRAWL_API_KEY`	Plan-safe
`skill`	Load one enabled skill’s `SKILL.md` or a contained resource.	Conditional · enabled skill exists	Plan-safe
`mcp_search_tools`	Search metadata for tools on configured MCP servers.	Conditional · connected MCP server	Denied
`mcp_select_tool`	Select one searched MCP tool; its full schema appears next turn.	Conditional · connected MCP server	Denied
`mcp_features`	Read MCP resources, invoke prompts, and complete arguments.	Conditional · connected MCP server	Denied
`mcp__<server>__<tool>`	Invoke one selected remote MCP tool.	Conditional · selected tool	Denied

Web fetch uses the strict URL policy. MCP servers are not a privileged bypass: once selected, their remote tools go through ordinary dispatch, approval, cancellation, and per-server concurrency controls.

Extensions running around the tools

Extensions subscribe to lifecycle seams around the agent loop. They are not tools the model invokes directly; they shape execution, visibility, persistence, and safety.

Extension	Current CLI behavior	Default
`EventStream`	Feeds typed model/tool lifecycle events into the terminal activity UI or headless NDJSON stream.	Always on
`PlanGate`	Enforces the plan-mode allowlist and the `docs/plan/` write fence, plus the parent orchestrate-mode allowlist, before approval is considered. Workers inherit plan restrictions live but treat orchestrate as normal mode.	Always on
`AutoApproval`	In auto mode, locally clears guarded native workspace mutations, then reviews remaining final tool input against the root request.	Active only in auto mode
`Approval` / `HeadlessGate`	Requests interactive approval or blocks action-capable headless calls unless auto-approved; auto delegates unresolved calls to `AutoApproval`, while yolo bypasses review.	Always on when applicable
`SessionHandler`	Records per-workspace JSONL sessions and restores them when requested.	On unless `--no-session`
`LongSession`	Compacts near the active context limit so interactive sessions can continue.	On; toggleable
`Truncation`	Caps tool output at 16,000 characters; the original is retained for result paging.	On; toggleable
`ToolRetry`	Retries eligible failures up to three times with 250ms backoff; deterministic native file-mutation errors are excluded.	Off; toggleable
`UsageMeter`	Accumulates model-reported token usage for the run.	Used by the host

› /extensions
› /extensions enable retry
› /extensions disable truncation

Toggleable extension choices persist in configuration and apply after the agent is rebuilt for the next run. Event streaming, safety gates, and the active UI wiring are intentionally not toggleable because disabling them would break the host contract.

What this CLI does not register

The workspace also includes an fs_admin_tools library bundle—copy_file, rename_file, delete_file, create_folder, and file_info—for restricted shell-less hosts. Orcacode does not register this bundle today. File information in the terminal agent comes from normal reading/searching, and filesystem administration remains available through the gated shell where appropriate.

Start here

Getting started

Work safely

Configure

Use the agent

Extend

Embed

Automate

Under the hood

Reference

Tools available now

Read this reference

Workspace files and search

Execution and compute

Coordination, memory, and recovery

Web and remote capabilities

Extensions running around the tools

What this CLI does not register

​Read this reference

​Workspace files and search

​Execution and compute

​Coordination, memory, and recovery

​Web and remote capabilities

​Extensions running around the tools

​What this CLI does not register

Read this reference

Workspace files and search

Execution and compute

Coordination, memory, and recovery

Web and remote capabilities

Extensions running around the tools

What this CLI does not register