Alternatives guide · open source
Open-source and source-available coding agents: choose the client, model, and privacy boundary
Open-source coding agents are not one category: compare the client licence, editor or terminal surface, connected model, tool permissions, and whether inference stays local before you choose.
What actually matters
- Separate an open-source client from open model weights. Cline, Kilo, OpenCode, Aider, and Continue can still send code to a hosted provider when configured that way.
- Choose the surface first: editor agents reduce context switching, while terminal clients make commands, diffs, and approval boundaries easier to inspect.
- For local inference, validate the endpoint, context length, quantisation, tool-call behaviour, and VRAM floor together rather than relying on the model name.
- BYOK shifts billing and retention to the provider you connect; a local server changes the privacy boundary but adds hardware and operations work.
- Check the current maintenance and service status of every client, including Continue, before standardising; test Cline, Kilo, or OpenCode if you need a supported agent loop.
The shortlist
- Cline for an open-source, editor-first agent with provider and local-model options.
- Kilo Code for editor, CLI, and native JetBrains paths with BYOK or hosted choices.
- OpenCode for an open-source terminal workflow with explicit provider connections and Plan mode.
- Aider for git-native pair programming and reviewable edits, or Continue for documented Chat/Plan/Agent workflows when the final unmaintained client and provider configuration fit your needs.
Evidence-backed entries
A sensible test workflow
- 01Decide whether the requirement is inspectable client code, local inference, provider independence, or simply lower cost.
- 02Start with a read-only repository map, then one documentation or test change with tool approvals visible.
- 03Repeat the task against the intended local or hosted endpoint and compare accepted patches, retries, latency, and spend.
Common mistakes
- Calling a BYOK client private without checking the provider's retention and training terms.
- Buying a GPU before measuring context length, concurrency, and the actual model quantisation.
- Expecting a small local model to reproduce a frontier agent loop without a scaffold-first evaluation.
Popular-AI shortlist snapshot: September 26, 2026. This guide is a starting framework, not a permanent ranking. Model prices, access, policies, and capabilities move quickly. Check the provider before committing spend or sending sensitive data.