Many models. One workbench. Your machine.
Run Claude, Codex, DeepSeek, GLM and more — side by side, or as a team. The open-source desktop workbench that turns many LLMs into one workforce you control.
Website · Download · Report an issue
English · 简体中文
Most AI coding tools hand you one agent, in one editor, on one folder, from one vendor. That's fine — until the job is bigger than one agent, the quota runs out mid-afternoon, or you need to know exactly what it changed before you keep it.
AgentLoom starts from the opposite assumption: many projects, many models, several agents working at once — on your machine, under your control.
-
One agent is a bottleneck. Run a team. Crown a lead — Claude or Codex, the models that are good at planning — and give it a bench of cheaper models. The lead splits the goal, hands out tasks in parallel, reviews what comes back and fixes what doesn't fit. Several files move at once instead of one, and you pay top-tier prices only for the thinking.
-
Never stuck behind one vendor. Out of Claude quota at 4pm? Point the same session at GLM, DeepSeek or a local model and keep going — same thread, same context, nothing to copy-paste. Switching vendors is a dropdown, not a migration. And when a session gets long, AgentLoom writes the hand-off brief so the next one starts warm.
-
Cheap models that actually finish the job. AgentLoom ships its own agent engine,
myagent, written in Rust — no Claude Code, no Codex, no vendor CLI required. Providers you connect through that engine get the same tool loop, plan mode and checkpoints that the expensive CLIs have, so a pay-as-you-go key gets a real shot at real work.
We measured that claim rather than asserting it — see Benchmarks below. -
See every move. Undo any of them. Every command and every file write lands as a card you can open, plus a Review panel with a file-level ledger. Keep what you like, roll back the rest, file by file. You don't have to be a terminal expert to tell whether the agent did the right thing — and you don't have to trust it blind to let it work.
-
Yours, truly. Open source and local-first. Your API keys live in your OS keychain, your conversations in a database on your own disk, and agents work directly in your own repositories. AgentLoom runs no server on the internet; the only listeners it starts are loopback-only helpers — the only things that ever leave your machine are the requests you send to the model provider you chose, and, if you turn on web search, your queries to the search backend you configured. Its own bookkeeping never touches your working tree.
Median resolved on a 30-instance subset of SWE-bench Verified.
myagentdrivingdeepseek-v4-pro, graded by the official SWE-bench Docker harness. Eight runs, range 16–19.
We quote the median, not our best run — a single run of this subset carries ±3 instances of noise, so any one number (including our 19/30) would overstate what you should expect.
| Engine · model | myagent (this repository, Rust) · deepseek-v4-pro, temperature 0 |
| Grading | Official swebench.harness.run_evaluation in Docker |
| Test leakage | None. The engine got a working test environment, but never saw the FAIL_TO_PASS grading test, and test.patch was never applied during a run |
| Runs | 8 graded runs · median 17 · mean 17.1 · range 16–19 (53.3%–63.3%) |
| Caveat on those runs | They span four days of engine iteration — replicates of a moving target, not eight repeats of one frozen build |
| Cost | roughly $3–6 in model spend for all 30 instances |
Read this before comparing: this is a hand-picked 30-instance subset, not the full 500 — it is not comparable to the public SWE-bench Verified leaderboard, and it deliberately excludes repositories with heavy C extensions our host could not build reliably. Frontier-model agents score 60–70%+ on the full set. The point of this number is "a mid-tier, pay-as-you-go model is enough for everyday work" — not "we beat frontier agents."
We publish the exact instance IDs and the full method, limitations included: docs/benchmarks.md · evals/swebench/fair30_ids.json. If you run the same 30 instances and get something materially different, please open an issue.
| Switch models mid-conversation | Diagrams and charts, rendered inline | Cards, review & undo, file by file |
|---|---|---|
 |
 |
 |
A session is the unit of work: one focused conversation that ships code, with checkpoints and undo. The sidebar holds every project you work on and every session inside it — no tab soup.
- Agent teams — configure any number of agents across providers (native CLIs like Claude Code and Codex, plus engine-driven providers like DeepSeek, GLM and Kimi); crown a lead, toggle members, dispatch work, watch results land.
- Session-centric workbench — all your projects in one window (GitHub repos and plain local folders alike), each with its own session list, groups, ⌘K search and hand-offs.
- Checkpoints & undo — a file-level write ledger with reviewable, selective undo. See exactly what an agent touched before you decide to keep it.
- Rich transcript rendering — mermaid diagrams, inline images, diffs, collapsible thinking, tool-call cards; long output folds by default so the conversation stays readable.
- Built-in agent engine (
myagent) — a Rust harness that runs provider-agnostic agent loops with tool use, plan mode, checkpoints and event streaming. Use it standalone on the command line, or let AgentLoom drive it. - Web search for every model — not every model ships with search; AgentLoom wires up third-party backends (DuckDuckGo with zero config, Brave/Exa with your key) so any agent can look things up.
- Bring your own everything — OpenAI-compatible and Anthropic-compatible endpoints, custom base URLs, local models.
- i18n — English and 简体中文 in the UI today, more on the way.
- Cross-platform — macOS (Apple silicon and Intel), signed and notarized. Windows is available as an unsigned preview until we can sign it. Linux is not started.
- macOS — grab the
.dmgfor your Mac (Apple silicon or Intel) from myagenthubs.com/agentloom or the Releases page, and drag it to Applications. Signed and notarized by Apple; it opens with no warnings. The website also serves a mirror, which helps if GitHub is slow where you are. - Windows — an unsigned preview is available: v0.1.0-windows-preview. It is a pre-release on purpose. We have no code-signing certificate, so SmartScreen will warn on it, and nobody has run it on real Windows hardware yet — CI compiles, packages and smoke-tests it, and that is all. Try it if a warning dialog does not put you off, and tell us what breaks.
Prerequisites: Rust (stable), Node.js ≥ 20, npm. Install the Tauri system prerequisites for your platform: https://tauri.app/start/prerequisites/
# 1. Build the myagent engine sidecar
cd harness-agent
cargo build --release
# 2. Place the sidecar where the app expects it (the command detects your platform's
# target triple)
mkdir -p ../app/src-tauri/binaries
cp target/release/myagent ../app/src-tauri/binaries/myagent-$(rustc -vV | sed -n 's/^host: //p')
# 3. Build / run the app
cd ../app
npm install
npm run tauri dev # development
npm run tauri build # release bundleapp/ Tauri desktop app (React + TypeScript front end, Rust back end)
harness-agent/ myagent — the provider-agnostic agent engine CLI (Rust)
docs/ Benchmarks, screenshots
evals/ Public benchmark instance lists
.github/ Issue and pull request templates, CI workflows
AGENTS.md Contribution rules for AI agents
- A signed Windows installer (the build already works; the certificate does not exist yet)
- More providers and deeper local-model support
- Richer agent-team workflows (discussion / round-table modes)
- GitLab support through the existing adapter seam
Issues, bug reports and pull requests are welcome — see CONTRIBUTING.md for how to build, test and submit. Contributors sign a lightweight CLA (one click in the PR) the first time they contribute.
AGPL-3.0. In short: use it freely, self-host it, fork it — but if you distribute a modified version or run one as a service, your changes have to be open too. That keeps the workbench honest for everyone.
The AgentLoom and MyAgentHubs names and logos are trademarks of MyAgentHubs and are not covered by the code license — see TRADEMARK.md.
© 2026 MyAgentHubs