Text tool protocol
The agent speaks through a single ```tool JSON block per turn — no proprietary function-calling. No tool block means the task is done. Runs on any OpenAI-compatible endpoint.
INCIPIENT is a general-purpose coding agent that works through a transparent tool protocol: every action is visible, every change waits for your approval, and everything is recorded in an audit log. Try the agent loop yourself right on this page — no API key, and none of your real files are touched.
Give the agent a task, watch the tool blocks it sends, then decide for yourself: allow or reject. Switch the approval mode to feel how its safety rules change — including the sandbox that rejects paths outside the workspace and the automatic block on destructive commands.
create file X containing … · read X · search for … ·
list files · run <command> · remember: … ·
/mode ask|auto|readonly · /plan <task> · /memory
Not a chatbot bolted onto an editor — INCIPIENT is an agent with a protocol, a permission gate, and an audit trail that anyone can inspect.
The agent speaks through a single ```tool JSON block per turn — no proprietary function-calling. No tool block means the task is done. Runs on any OpenAI-compatible endpoint.
read_file, list_dir, grep, write_file, edit_file, run_command, and remember — enough for a real read → change → verify cycle.
Reads are always free. Writes and commands ask for permission in ask mode, run immediately in auto — except destructive ones — and are always rejected in readonly.
Every path is normalized and confined to the project folder. Escape attempts — even ../../etc/passwd — are rejected before the tool ever runs.
Important facts live in .incipient/memory.md via the remember tool — read back at the next session, so the agent never starts from zero.
Every tool call — allowed, rejected, or blocked — is recorded with a timestamp in .incipient/audit.log. No silent actions.
The exact same agent loop you are running in the console above.
The model takes your task and replies with plain text — or a single ```tool block holding the tool name and its JSON arguments. One turn, one action.
Before anything runs, the gate checks the approval mode, the sandbox path, and the destructive-command list. This is where you press Allow or Reject.
Tool output — file contents, grep results, test results — goes back as context for the next turn and lands in the audit log. The loop stops when the model replies without a tool block.
This matrix is not theory — switch modes in the console and prove it on the same task.
| Action type | ask | auto | readonly |
|---|---|---|---|
| Read file · list · grep | Free | Free | Free |
| Write / edit file | Asks first | Runs immediately | Rejected |
| Ordinary shell command | Asks first | Runs immediately | Rejected |
| Destructive command (rm -rf, mkfs, curl | sh…) | Asked + warning | Blocked outright | Rejected |
| Path outside the workspace | Sandbox-rejected | Sandbox-rejected | Sandbox-rejected |
This page simulates the model. The real INCIPIENT is a TypeScript CLI + MCP server that connects the same loop to a real model through any OpenAI-compatible endpoint.
ask / auto / readonly modes, sandbox, and audit log you just tried.# check that your environment is ready $ incipient doctor # give it a task — ask mode is the default $ incipient run "add a /health endpoint and its test" # or run it as an MCP server for other clients $ incipient mcp