One binary, no runtime
Go builds it; the terminal runs it. Classic TUI out of the box, an optional opentui interface when you want the framed inspector and live preview.
Wombat
A general-purpose agent. In your terminal. Finishes the job.
Edit, run, verify, browse and research with one binary that keeps its tool registry small and its research bounded — then hands back when the work is done, not when it runs out of ideas.
curl -fsSL https://raw.githubusercontent.com/ramsterr/Wombat/main/install.sh | bash 18/18 tasks passed 2.02× faster than Pi on the suite 41% fewer prompt tokens
benchmarks
Faster than Pi across a full suite, and cheaper on tokens.
2.02× faster total wall time
133.6s vs 270.2s, 18 tasks
133.6s vs 270.2s, 18 tasks
Same model, same tasks, same skill and MCP server, fresh workspace per run. Wombat passes 18/18 with the lean profile, uses 41% fewer prompt tokens, and finishes the research task Pi times out on. The comparison page also shows where Pi is still the lighter harness — small, single-file work.
The parts that make it different
Not a feature list for its own sake: each of these changes how a run behaves, and each one is documented with the exact commands in the docs.
Nine tools by default
The coding profile exposes read, write, edit, bash, glob, grep, apply_patch, tool_search and request_capability. Web, LSP, verify and media load mid-run when the task needs them.
Research with a budget
Quick gives 2 searches and 1 fetch, standard 5 and 3, deep 10 and 6 — each with a hard ceiling. When the budget is spent the model is told to synthesize, not to keep looping.
A fleet when you ask
Plan a ledger, run scouts and writers in isolated git worktrees, pass reports down the dependency graph, and merge only what passes its checks.
Sessions and checkpoints
JSONL transcripts per project, resume by id or last, and shadow-git snapshots so a bad edit is one command from gone.
MCP, skills, custom agents
stdio and HTTP servers with OAuth and per-server gates, skills as markdown, and specialists with their own prompt, tools, permissions and model.
quick start
From clone to first task in two commands.
The installer gives you the latest tagged release. If you want the capability profiles,
research budgets and custom agents described on this site, install from
main — the tag is behind the branch.
# current main
go install github.com/ramsterr/Wombat/cmd/wombat@main
wombat setup # provider, key, model
wombat doctor # one small request, errors surfaced
wombat # open the chat
# or the tagged release
curl -fsSL https://raw.githubusercontent.com/ramsterr/Wombat/main/install.sh | bash Optional opentui interface.
cd tui && npm install && npm run build
once (Node 26.4+), then run Wombat from that checkout. Without it you get the classic
TUI, which has every feature except the rendering. Switch back with
wombat classic.
- runtime
- none — one static Go binary
- providers
- OpenAI-compatible, Anthropic, OpenCode, local servers
- default profile
- 9 tools, capabilities on demand
- research
- off · quick · standard · deep, with hard ceilings
- fleet
- 5 scouts · 2 writers · git worktrees
- state
- ~/.wombat — config, sessions, checkpoints, agents
A fleet, only when the work is parallel
The lead plans, workers execute, checks decide. Nothing about the fleet is implicit: a one-step plan is a single agent.
Dependency graph, not rounds
Free slots start newly-ready work, long chains are scheduled first, and a slow scout never holds up an independent writer.
Owned files
Writers declare the files they may change. Writes outside that set are refused, which is what makes automatic merges safe.
Worktrees and checks
Work happens in a git worktree; checks run once, review policy decides who reads the patch, and the merge is three-way.
Resumable workers
A worker that needs a decision returns needs_input; resume_worker continues the same session and worktree.
Documentation
Eight guides, matched to main after v0.2.6, with the exact commands quoted
from the repository.
01
Getting started
Install, choose a provider, build the optional UI, and run a first task.
open → 02Tools & profiles
The nine-tool coding profile, capability loading, and research budgets.
open → 03Agents & orchestration
Custom agents, the plan scheduler, worktrees and review policy.
open → 04Wombat vs Pi
The benchmarks, the method, the caveats, and the raw data path.
open →install
One binary. One config file. Start with the docs, then decide about the fleet.