Redcast Tools, permissions and effort
What native agents can do, when they ask, and how a run ends.
A native agent exposes only the tool families enabled in its profile, plus MCP tools selected for the conversation. An ACP agent runs its own tools instead; the controls below describe Redcast's native agent loop.
Tool families
| Family | Typical tools | Permission behavior |
|---|---|---|
| Files | list_dir, read_file, search_files, write_file, edit_file | Reads run directly; changes ask by default |
| Terminal | run_command | Commands ask by default |
| Web | web_search, web_fetch | Follows Web permission and URL scope |
| Memory | memory_read, memory_write | The agent's own notes |
| Notes, opt-in | notes_search, notes_read, notes_append | Appending to your Notes asks, even with general changes allowed |
| Screen, opt-in | screen_context | Foreground user turns only; asks and names the app |
| Sub-agents, opt-in | delegate | Read-only delegated tasks within the parent's budget |
| Planning and skills | update_plan, load_skill | A plan for the run and instructions from allowed skills |
Document reading handles PDFs by page, scans through OCR, images and supported document formats. It is bounded: ask for specific pages or sections when a file is large. Search Files uses Spotlight for the agent's search tool; the launcher's separate File Search feature has its own private index.
Permission rules
Ask Each Chat asks before an operation unless you granted that family and risk for the conversation. Always Allow permits it, including unattended work. Never Allow refuses it and can remove that capability from the offered tools.
These are application checks, not a sandbox around a shell command. A command can affect whatever the launched process can access. Review the actual command and working directory in the tool row.
The agent's own memory is different from the user's Notes. Allowing memory writes does not allow appending to Notes without the Notes permission flow.
Effort budgets
| Effort | Tool rounds | Total tool output | Background time | Concurrent sub-agents |
|---|---|---|---|---|
| Quick | 8 | 256 KB | 5 minutes | None |
| Balanced, default | 25 | 512 KB | 15 minutes | Up to 2 |
| Thorough | 60 | 1 MB | 45 minutes | Up to 4 |
These are ceilings, not a target to fill. A simple question should end quickly. A multi-step task can keep a plan and use the available budget. Sub-agents require the tool family to be enabled; each has up to 12 rounds and 256 KB of output inside the parent's run.
Agent effort is separate from model reasoning effort. Changing one does not automatically set the other.
At a limit
The agent is asked to wrap up before its budget ends. At the limit tools are disabled and it hands over what it finished, what remains and a next step. A background run also has a deadline. A limit should not silently discard work already done.
Long tool results and old conversation context may be condensed to fit the model's window. If exact wording matters, tell the agent to read the relevant source again and cite the file or page.
Stop, queue and steer
Stop cancels the run and its running command, retry waits and delegated tasks. A permission answer arriving after cancellation cannot revive the stopped operation.
Return while a reply runs queues the next message. Send Now (⌘Return) steers a native agent at its next step. A call already in progress may finish before it reaches that step. Other provider routes stop their reply before sending immediately.
Retries and sources
The loop can retry temporary network and rate-limit failures, and trim older context after an overflow. It reports retry activity. A persistent credential, balance or unsupported-model error needs a change from you.
Web pages and tool output are supplied as material to inspect. Their contents do not grant permission to run commands. The web guide explains the narrower URL scope used after private data has been read and in background runs.