MomongaCode

Features

  • Terminal UI and headless mode. An interactive session, or -p/--print with --output-format text|json|stream-json.

  • Models. Connections to Ollama, chat-completions servers, Open Responses servers, Anthropic, and server programs that MomongaCode starts itself, such as Apple's fm. A model is called by the name its server gives it. models.fallback lists the models tried next when a model cannot answer. momonga gateway probe <model> measures a model's capabilities and records the result.

  • Built-in tools. Bash (with BashOutput and KillBash for background work), Read, Write, Edit, Glob, Grep, ReadOutput, UpdatePlan, MemorySave, MemorySearch, HistorySearch, HistoryRead, RulePropose, Agent (subagents: the general subagent general-purpose is built in and needs no setup), LoadTools, and activate_skill, and the operations tools for Git, pull requests and issues (GitStage, GitCommit, GitPush and the others): each states what it changes, and the approval names it. The reference lists every tool.

  • Permission modes. The mode decides what a call that no rule decides does:

    • manual asks before every call that changes something or reaches the network;
    • acceptEdits also allows, without asking, edits and file-system commands inside the workspace;
    • plan refuses every tool that writes or changes something and asks before a command that is not read-only, so the model reads and proposes a plan;
    • auto lets a classifier model decide each such call in your place, as the next item says;
    • bypassPermissions allows every call that no deny rule matches and asks nobody.

    In every mode a deny rule still denies, the secret locations stay denied (see Permission rules), and a command that removes the workspace is asked of you except in bypassPermissions. Rules are written as Tool(specifier) and set in permissions.allow, permissions.ask, and permissions.deny (Configuration). momonga approve lists and resolves pending approvals with allow_once, allow_always, reject_once, or reject_always.

  • Policy verification. momonga policy verify asks the daemon to check the rules of the folder: that no allow rule gets around a deny rule, that no rule leaves a route to secrets, and that every rule can take effect. It states what is proved and what is not, and lists a rule that is redundant, shadowed by another, or cannot be reached as a note; a note refuses nothing. The verdict is stored with the rules of the folder, and a run of that folder reuses it until a rule changes.

  • Verifiable completion. A run can declare completion conditions: shell commands that MomongaCode runs itself, in the sandbox, when the model stops. A run ends COMPLETED only when every condition it owes passed on the final state of the workspace; otherwise the failures go back to the model, and a run that stops again without a change ends STOPPED with the reason conditions_unmet and names the conditions. momonga report shows the outcome (see Usage).

  • Auto mode. momonga auto-mode measures how well a model judges approvals for auto mode. Auto mode is honored only from user settings or the command line.

  • Sandbox. Bash runs under Seatbelt on macOS and bubblewrap on Linux. Windows has no sandbox in this release: Bash asks before every command.

  • Memory and rules. momonga memory and momonga rule (see Why MomongaCode?).

  • Skills and agents. Skills are SKILL.md files found in .momonga/skills and .agents/skills of the project and in ~/.momonga/skills and ~/.agents/skills. A general subagent, general-purpose, is built in; custom agents are read from .momonga/agents directories, and a custom agent named general-purpose replaces the built-in one. Project skills are loaded only for trusted folders.

  • MCP. Servers over stdio or HTTP with OAuth. momonga mcp list and momonga mcp login <server>. Pending MCP input requests are handled with momonga elicit.

  • Hooks. Commands run on PreToolUse, PostToolUse, UserPromptSubmit, Stop, and SessionStart. A hook that exits with code 2 blocks.

  • Editor integration. momonga acp speaks the Agent Client Protocol on standard input and output.

  • Observability. momonga trace shows the model and tool spans of a run, momonga cost shows what it cost, and an OTLP endpoint can be configured under telemetry.otlp.

  • Background jobs and runs. momonga status, momonga jobs, momonga attach, and momonga cancel.

  • Delegated work. A mandate hands work over with bounds, as Usage describes: --mandate <name> on momonga -p, or /delegate <mandate> <request> in the interactive session.

  • Measured reliability. momonga reliability reports how runs ended, momonga eval bench <folder> runs a fixed bench set on one model, and momonga permissions review samples recorded permission decisions for review (Usage).