Open source · AGPL-3.0 · Self-hosted · Built in Go

The open source harness.Many agents, any model.

TaskWing turns one coding task into a team of AI agents, each working in its own sandbox, on any model you choose. You get back one pull request, with tests that prove it works.

your-org/shop
  • boatWorks today
  • E2BPlanned
  • AWSPlanned
  • DockerPlanned, runs on your machine
  • MixedClaude plans, open models build
  • ClaudeEvery step on Claude
  • OpenAIEvery step on GPT
  • Open-weightQwen, DeepSeek and more

Or in your terminal: taskwing run "add discount codes at checkout, with tests"

What you get

A team of agents, one result you can trust.

A team, not a chat

One task in.
One tested pull request out.

TaskWing splits your task into pieces, gives each piece to an agent in its own sandbox, and joins the work on one branch. You approve the checks, and you merge.

Hidden exam

Workers never see the tests, and no agent grades itself. The exam decides.

Its own sandbox

Code goes in as a git bundle, and only commits come out. No GitHub login ever enters a box, and each box is scanned for secrets before it is deleted.

Open source

AGPL-3.0.
Read every line.

26 of 67 features built · Nov 2026

Any model

Claude now · GPT and open models next

You merge

Every run ends as a draft pull request

How a run works

Four steps. You decide twice.

You approve the exam before any agent starts, and you merge the pull request at the end. Everything in between runs on its own.

01You approve the checks

Plan

Your coding agent splits the task into pieces that touch different files, and pins the public names so they fit.

02Agents in parallel

Build

Each piece gets a fresh sandbox and an agent. Code goes in as a git bundle, and only commits come out.

03No agent grades itself

Exam

The pieces are joined on one branch. Hidden tests and every existing test run in a clean sandbox.

04You merge

Pull request

Your machine opens one draft pull request with the proof and the exam results. TaskWing never merges.

The hidden exam

The agent said done. The exam decides.

Agents that can see the tests learn to pass the tests. So a separate session writes the exam, and the tests stay out of every sandbox until the work is finished.

Who sees what

Youapprove
the checks, in plain words
Leadplans
the checkstests hidden
sandbox wall
Workersbuild
their piecethe namestests hidden
Exam boxgrades
joined workhidden testsold tests

You read plain sentences

"A retry after a 500 returns 200." You approve the checks, not the code.

Workers never see the tests

Neither does the lead. Only the clean exam box gets them.

Old tests count too

A piece that breaks something that used to work fails the run.

The open harness

Any model. Any sandbox. Your machine in charge.

TaskWing owns the part in the middle: splitting the work, running it in parallel, checking it, and handing it back. Models and sandboxes plug in.

TaskWingGo engine on your machine
Plan with a frontier modelThe split and the exam need the best reasoning you have.
Build routine work on open modelsMost agent steps are routine. Send them where they cost less.
Grade with hidden testsWhichever model built it, the same exam decides.
ModelsWhat agents think with
Claude (works now)GPT (next)Open-weight models (next)
AgentsThe coding CLI in each box
Claude Code (works now)OpenCode (next)Codex CLI (next)
SandboxesWhere agents run
boat (works now)Local folders (works now)E2B (next)AWS Nitro (next)Docker (next)
Real runs

Every run leaves a receipt.

Time, checks, pieces, and the pull request. These are from the first runs of the proof of concept in November 2026.

Hidden exam · first 3 runs

12/ 13

Hidden checks passed, graded in a clean sandbox that no worker could see into. Every check is written before any agent starts.

Fastest run

1m 20s

Task to tested pull request

4

Bugs proven

0 merge conflicts2,076 lines from 2 agents on one branch
build run · test repository

Add a health endpoint with a test

Agents
2
Sandboxes
boat
Secrets in boxes
none
Time
1m 20s
Exam4 / 4 passed
bug hunt · Osprey

Find and fix bugs in a Go service

Bugs proven by a failing test
4
Fixers
4
Time
8m 58s
Pull request
#120 merged
Exam5 / 5 passed
build run · Osprey

Evaluation list, dispositions, alert webhook

Pieces
2
Merge conflicts
0
Lines added
2,076
Time
15m 12s
Exam3 / 4 passed

The 3 of 4 run failed on a field name the exam guessed. The lead now pins public names before the exam is written, and that failure is why.

Bug hunt

No failing test, no bug.

Point TaskWing at a part of your code. It writes a test for every bug it suspects and keeps only the tests that fail on today's code.

Suspect Failing test Fixer + exam
Questions

What is TaskWing, in one table?

At a glance
TaskWing at a glance
LicenceAGPL-3.0, free
EngineGo, runs on your machine
Desktop appmacOS, Linux, Windows · demo engine for now
Sandboxesboat today · Docker, E2B and AWS planned
ModelsClaude today · GPT and open models next
Each run ends as1 draft pull request · you merge
StatusProof of concept, November 2026 · 26 of 67 features built
Every feature and its status ↗

What is TaskWing?

TaskWing turns one coding task into a team of AI agents, each working in its own sandbox. You approve the exam first, and you get back one pull request, with tests that prove it works.

Is TaskWing free?

Yes. TaskWing costs $0 and is open source under the AGPL-3.0 licence. You bring your own agent and sandbox accounts, such as Claude Code and boat. The full source is on GitHub.

Where do the agents run?

Each agent runs in its own sandbox: boat today, with Docker, E2B and AWS planned. Code goes in as a git bundle and only commits come out. No GitHub login ever enters a box.

Which models can it use?

The lead is your own coding agent, Claude Code today. GPT and open models are next, with each step routed to the model that suits it.

Does TaskWing merge code for me?

No, never. A run goes no further than 1 draft pull request with the proof and the exam results. You review it and you merge.

What does the desktop app run on?

3 platforms: macOS (one universal build for Apple silicon and Intel), Linux and Windows. The Mac test build is about 4 MB. It runs on a demo engine until the engine stream ships.

Open source · AGPL-3.0

Give TaskWing one task.

A team of AI agents works on it in parallel, and you get back one tested pull request. Your machine, your models, your merge.

GoSelf-hostedBring your own modelsDraft PRs only