⌘
App developers
Build web, mobile and API features end to end. Agents test your UI in a real browser and hand back reviewed code.
"Add Stripe checkout with webhooks and tests"
v1.0 for Windows · free with your own models
One prompt.
A whole engineering team.
Other coding agents give you one assistant. SwarmCode gives you a team: a planner, parallel builders, a tester, a reviewer and a security check. You get back one clean diff with proof that it works.
CLI · any OS
npm i -g github:yogeshwardev/MYCODE
Windows app 101 MB · Windows 10/11 · open source, read the code
Leancontext: journals and recall instead of re-reading
10/12real tasks passed, graded by hidden tests (SwarmCode · Agent)
Anymodel: cloud, local or free
35expert skills, picked for you
Works with Claude · GPT · Gemini · DeepSeek · OpenRouter · Ollama and any OpenAI-compatible model
How it works
From request to reviewed diff, automatically.
Other tools give you parallel agents and leave the splitting, merging and checking to you. SwarmCode runs the whole loop.
- 1PlanA planner maps the code and splits the work into a dependency graph with clear file ownership.
- 2Build in parallelAgents work at the same time, each in its own isolated git worktree.
- 3MergeResults merge as they finish; an agent resolves any conflicts.
- 4TestYour test command runs; failures go back to the agents automatically.
- 5ReviewAn independent reviewer and a security gate check the combined change.
- 6ApplyYou get one diff and a proof report. Apply it, or keep it on a branch.
Built for
Whatever you build, it has a playbook.
¶
Docs and data
Read PDFs, Word and Excel files, analyse data, and generate polished PDF reports straight from your project.
"Turn these 3 PDFs into a one-page summary PDF"
◈
Security teams
Reviews, secure fixes, threat models, dependency audits, cloud hardening and incident write-ups, with a security gate on every change.
"Audit this repo and fix every high-risk finding"
Features
Built to be trusted, not just fast.
Parallel agents that finish the job
Big tasks are planned, split and built by several agents at once, then merged, tested and reviewed. Small tasks go straight to a single agent. SwarmCode picks, and asks before spending more.
Hidden-test guarantee
Before any code is written, a separate agent turns your request into acceptance tests the builders never see. The work is done only when they pass.
Best of 3
Three independent attempts, even on different models. Each is graded on your tests and on security, and a judge picks the winner.
Proof for every change
A report with every check, the files changed, models, tokens and cost. One command turns it into a pull request.
Security gate
New secrets, injection risks or disabled TLS that a change introduces are caught and fixed before you see it.
Any model, your keys
Claude, GPT, Gemini, DeepSeek, OpenRouter or free local models. A strong model plans while fast ones build.
Long sessions without amnesia
Every agent keeps a searchable journal, so it can recall file contents and decisions even after its context is compacted. Huge logs are kept out of context and read on demand, which cuts token use.
Time travel
Every request is checkpointed. Replay what happened, or re-run any request from its exact starting code with another model.
Health radar
One scan for secrets, risky code, vulnerable dependencies, failing tests and hotspots, each with a one-click fix.
Browser, PDFs and documents
Agents test your web app in a real browser, read PDFs and Word files, and produce polished PDF reports.
Compare
What you get that others don't.
Every agent can write code. SwarmCode is built to finish the job and prove it.
| SwarmCode | Claude Code | Codex CLI | opencode | Cursor | |
|---|---|---|---|---|---|
| Splits one task across parallel agents, then merges, tests and reviews automatically | ✓ | ◐ Subagents you direct | ◐ Parallel tasks, you merge | ◐ Subagents you direct | ◐ Parallel agents, you pick |
| Hidden acceptance tests written before the code | ✓ | ✗ not built in | ✗ not built in | ✗ not built in | ✗ not built in |
| Independent reviewer and security gate before you see the diff | ✓ | ◐ /security-review and hooks, set up by you | ◐ /review, run by you | ◐ Custom commands | ◐ Bugbot, a separate product |
| Best of 3 across different models, graded automatically | ✓ | ✗ not built in | ◐ Multiple attempts in the cloud, you pick | ✗ not built in | ◐ Compare models, you pick |
| Proof report with every check, cost and token count | ✓ | ✗ not built in | ✗ not built in | ✗ not built in | ✗ not built in |
| Repo health scan with one-click fixes | ✓ | ✗ not built in | ✗ not built in | ✗ not built in | ✗ not built in |
| Works with any model and your own keys | ✓ | ◐ Claude models; others via gateways | ◐ OpenAI models; local via --oss | ✓ | ◐ Its supported models |
| Free local models (Ollama) | ✓ | ✗ not built in | ✓ | ✓ | ✗ not built in |
| Open source | ✓ | ✗ not built in | ✓ | ✓ | ✗ not built in |
| Price | Free + your model | ◐ Subscription or API | ◐ ChatGPT plan or API | Free + your model | ◐ Subscription |
✓ built in · ◐ partly available: needs manual steps, setup or a separate product · ✗ not built in (you could still ask the agent to do it by hand). Column headers link to each product's documentation; checked October 2026. Products change quickly, so tell us if something is out of date and we'll fix it.
The app
Calm enough to work in all day.
Benchmarks
Measured, not claimed.
Twelve realistic tasks, from small bug fixes to multi-feature builds, graded by hidden acceptance tests the agents never see. Every tool ran the same model.
Tasks passed, same model (gpt-oss:120b via Ollama Cloud): SwarmCode · Agent 10/12 · SwarmCode · Swarm 9/12 · Codex CLI 11/12 · opencode 6/12.
SwarmCode · Agentours
10 / 12 passed
- Avg time
- 1m 40s
- Median tokens
- 172k
- Small tasks
- 3/4
- Medium tasks
- 3/4
- Large tasks
- 4/4
SwarmCode · Swarmours
9 / 12 passed
- Avg time
- 4m 21s
- Median tokens
- 387k
- Small tasks
- 4/4
- Medium tasks
- 2/4
- Large tasks
- 3/4
Codex CLI
11 / 12 passed
- Avg time
- 2m 53s
- Median tokens
- 442k
- Small tasks
- 4/4
- Medium tasks
- 4/4
- Large tasks
- 3/4
opencode
6 / 12 passed
- Avg time
- 1m 2s
- Median tokens
- not reported
- Small tasks
- 4/4
- Medium tasks
- 1/4
- Large tasks
- 1/4
| Task | SwarmCode · Agent | SwarmCode · Swarm | Codex CLI | opencode |
|---|---|---|---|---|
| Fix a date bugsmall | ✓ pass10s · 29k tok | ✓ pass27s · 32k tok | ✓ pass1m 57s · 158k tok | ✓ pass33s |
| Fix cents roundingsmall | ✓ pass51s · 119k tok | ✓ pass1m 8s · 78k tok | ✓ pass1m 38s · 254k tok | ✓ pass2m 6s |
| Email validationsmall | ✗ fail1m 28s · 174k tok | ✓ pass2m 55s · 205k tok | ✓ pass2m 11s · 458k tok | ✓ pass53s |
| Rename across codebasesmall | ✓ pass24s · 100k tok | ✓ pass1m 21s · 126k tok | ✓ pass2m 31s · 311k tok | ✓ pass1m 11s |
| Functions across 3 modulesmedium | ✓ pass1m 8s · 169k tok | ✓ pass3m 58s · 660k tok | ✓ pass2m 56s · 404k tok | ✗ fail37s |
| Pagination + callersmedium | ✓ pass2m 36s · 908k tok | ✓ pass5m 19s · 885k tok | ✓ pass2m 36s · 430k tok | ✓ pass1m 45s |
| CSV exportmedium | ✗ fail1m 8s · 144k tok | ✗ fail1m 25s · 94k tok | ✓ pass2m 47s · 490k tok | ✗ fail21s |
| Stock + cancellationsmedium | ✓ pass1m 13s · 313k tok | ✗ fail4m 14s · 324k tok | ✓ pass3m 37s · 572k tok | ✗ fail24s |
| 4 store featureslarge | ✓ pass1m 20s · 309k tok | ✓ pass6m 19s · 757k tok | ✓ pass3m 7s · 500k tok | ✓ pass2m 12s |
| 4 data structureslarge | ✓ pass1m 39s · 313k tok | ✗ fail8m 16s · 1.4M tok | ✓ pass2m 30s · 400k tok | ✗ fail15s |
| 8 utility functionslarge | ✓ pass7m 22s · 2.2M tok | ✓ pass4m 43s · 451k tok | ✗ fail3m 30s · 454k tok | ✗ fail43s |
| 3-report modulelarge | ✓ pass43s · 133k tok | ✓ pass12m 9s · 1.9M tok | ✓ pass5m 17s · 1.2M tok | ✗ fail1m 22s |
What it costs: a typical medium task used about 241k tokens in agent mode. That is roughly $0.56 at Sonnet-class API prices before prompt caching (caching lowers it), and $0 with local models through Ollama. Swarm and best-of-3 modes use more; SwarmCode shows the estimate and asks before using them.
Model for every tool: gpt-oss:120b via Ollama Cloud. Run on October 15, 2026. One run per task and tool, same model for every tool, graded by hidden acceptance tests the agents never saw. Single runs vary, so treat small differences as noise. Token counts are shown where a tool reports them; opencode does not.
These results are from SwarmCode 1.0.0. Since then, agents trim old tool output from their context: re-running the two most token-heavy tasks in agent mode used 590k–879k tokens instead of 2.18M (8 utility functions) and 125k–234k instead of 313k (4 data structures), with the same pass results. A full re-run with several runs per task is next.
How it works: each task starts from a fresh copy of a small real-world project. When an agent finishes, hidden tests are copied in and run; pass means every requirement works. The harness is open source, so you can run it on your own models: npm run bench -- --external "name=your-cli {prompt}".
Skills
Expert playbooks for every kind of work.
SwarmCode picks the right ones for each request. Type / to choose them yourself, and add your team's own.
Build
new-project · web & frontend-quality · ui-design · mobile-app · api-design · database · llm-apps · migration-upgrade · refactoring · performance
Ship
test-writing · web-qa · debugging · code-review · git-workflow · devops-ci · deployment · docs-writing · explain-codebase
Work
documents (PDF, Word, Excel) · data-analysis · automation-scripts
Secure
security-review · secure-fix · threat-model · dependency-audit · secrets-hygiene · hardening · cloud-security · api-security-testing · compliance-evidence · detection-engineering · incident-forensics · suspicious-code-triage · pentest-report
Privacy
Your code stays yours.
- Runs on your machine. Code goes only to the model provider you choose, or nowhere with local models.
- API keys are encrypted with your OS keychain.
- No telemetry.
- Every edit and command needs your approval, unless you turn on autonomous mode.
- A full audit trail of what every agent did.
FAQ
Questions
Is it free?
Yes. SwarmCode is open source (Apache-2.0). You pay your model provider for usage, or use free and local models through Ollama.
Which models work best?
Any capable tool-calling model works. A strong model (such as Claude Opus or GPT-class) for planning and review, with a faster one for the parallel workers, gives the best balance of quality and cost.
Do I need git?
For parallel agents and best-of-3, yes: each agent works in its own git worktree. The single agent works in any folder.
Windows says the app is unrecognized.
Early builds are not code-signed yet. Click "More info", then "Run anyway". Signed builds are on the way.
macOS and Linux?
The CLI runs everywhere (npm i -g swarmcode). Desktop builds for macOS and Linux come from the same code and are coming soon.
Can my team share skills?
Yes. Turn a good session into a skill with one command, and share it with SHA-256 pinning and Ed25519 signatures so nobody can tamper with it.
Stop babysitting your AI.
Give it a team.
Free and open source. Bring any model, or run fully local.