GitHub

docs / overview

Documentation

Wombat is a general-purpose agent that runs in your terminal. One binary, no runtime, and a harness built around doing the whole job: edit, run, verify, browse, research, and hand back when it is done.

What ships in the box

The default experience is a chat UI with a framed inspector, a live footer and a command palette. Under it is a single agent loop with a small tool registry that grows only when the task needs more reach.

areawhat you get
EditingRead, write, edit and atomic multi-file patches, with exact-bytes handling for CRLF and UTF-8 files.
Shellbash with a 30s default timeout for foreground work, background jobs, and a guard that rejects searches outside the workspace.
Gitbash git. Dedicated git tools were removed, and checkpoints use a private shadow repo.
Verificationverify, the browser tool, and project checks; a writer's patch must pass checks before it merges.
WebSearch across several engines, page fetch with readable-text extraction, and image search — loaded as capabilities, not always-on.
ResearchQuick, standard and deep modes with soft budgets and a hard ceiling, plus one bounded extension request.
FleetLead plus scouts and writers in git worktrees, a dependency-graph plan scheduler, and resumable workers.
SessionsResumable JSONL transcripts scoped to a directory, checkpoints you can restore, and project references.
MCPLocal stdio and remote HTTP/SSE servers, per-server permission gates, and OAuth with PKCE and token refresh.
AgentsCustom specialists with their own prompt, tools, permissions, MCP servers, skills and model.

Start here

The shape of the harness

Three ideas explain most of the behaviour you will see, including the benchmark numbers on the comparison page.

1. Tools are a budget, not a feature list

Every exposed tool contributes its JSON schema to every request. When Wombat exposed 38–39 tools, roughly 71% of the first prompt was tool schemas. The default coding profile now exposes nine, and tool_search / request_capability let the model pull in web, semantic, verify or media tools mid-run. See tools & profiles.

2. Research is bounded

A research loop that never ends is worse than a slightly shallow answer. Quick mode allows two searches and one fetch, standard five and three, deep ten and six, each with a hard ceiling. When the soft budget is spent, the next call returns "synthesize now" instead of another page.

3. Parallel work is explicit

The lead plans a ledger, scouts and writers run in isolated worktrees, dependent steps receive upstream reports, and a patch must pass its checks before merge. Nothing about the fleet is implicit: if the plan has one step, the run is one agent. See agents & orchestration.

Version and status

The latest tag is v0.2.6; main carries the unreleased work covered here — custom agents, the skills browser, capability profiles, bounded research, nine-tool coding profile, and the rebuilt opentui interface. Where this site says "current", it means main after that tag.

wombat version
wombat doctor
wombat config