These are practical first configurations for common workflows. Treat them as a testable baseline, then measure successful work, review effort, and total cost.
Budget is a pilot hypothesis: compare retries, review, hosting and plan limits. A low-cost starting point for developers who need useful patches without paying frontier prices for every prompt.
Solo developers, students, and small teams with repeatable coding work.
Suggested stack
GPT-6 Luna or DeepSeek V4.1-Flash for routine generation, tests, and repository search.
Cursor for tight editor loops, or Claude Code when terminal delegation is the bigger need.
A local test suite and explicit review checkpoint before any broad change lands.
Tradeoffs
Cheaper models can spend more steps or need more steering on ambiguous tasks.
DeepSeek's open-weight path adds serving and policy work that a hosted subscription avoids.
A budget stack only saves money when retries and human review stay under control.
Check current product eligibility, plans and client terms before standardizing a beta tool. A margin-aware AI stack for agencies that need client separation, repeatable delivery, and a fallback path.
Small agencies and studios delivering content, research, design, or engineering work.
Suggested stack
Cursor for production teams that live in the editor, plus Claude Code or Muse Code for deeper terminal work.
A model gateway or documented provider matrix so every client does not become a bespoke integration.
Per-client data boundaries, approval rules, and a cost-per-deliverable dashboard.
Tradeoffs
Standardising the workflow matters more than forcing one model on every client.
Beta agents can be valuable pilots but should not become the only path to delivery.
Contractor access, data retention, and client terms can outweigh raw model quality.
Local inference alone does not certify privacy; review telemetry, extensions and network exposure. A self-hosted coding loop for developers who want agent help without routing every prompt through a cloud API — Continue or Cline in the editor, Ollama or MLX on the Mac, and optional GPU sizing when a desktop card beats Apple Silicon.
Solo developers, privacy-conscious builders, and hobbyists who accept local ops work in exchange for data control.
Suggested stack
Continue or Cline for the editor surface; both support agent workflows. Continue upstream is read-only/unmaintained, so evaluate provider compatibility, permissions and maintenance before adoption.
Ollama on port 11434 (or MLX on Apple Silicon, LM Studio on 1234, llama.cpp server on ~8080); choose quantisation for the hardware, then validate structured outputs and tool calls on the exact model/runtime pair.
For coding agents, current Ollama docs recommend at least 64K context where hardware allows. Set OLLAMA_CONTEXT_LENGTH or num_ctx, then use `ollama ps` and the harness prompt/logs to confirm effective context before wiring a workflow.
Optional hybrid: escalate ambiguous refactors to a usage-capped cloud model while routine edits stay local.
Tradeoffs
Local models need operator time for pulls, context tuning, and quant trade-offs; structured-output reliability varies by model, quantisation, and runtime, so validate it rather than assuming a universal quality floor.
Aider needs ollama_chat/model, not bare ollama/; OpenCode expects an OpenAI-compatible /v1 base URL.
A solo local stack only wins when review discipline stays intact; privacy does not remove the need to read diffs.
A fast-moving default for small teams: one primary agent surface (Cursor or Claude Code), a usage-aware cloud model tier, and an optional local fallback for bursts or privacy-sensitive slices.
Seed-stage and Series A engineering teams that need speed without letting agent spend or outages block delivery.
Suggested stack
Cursor for continuous editor work, or Claude Code when terminal delegation and long-horizon tasks dominate.
A documented model routing rule: frontier for ambiguous cross-system work, cheaper tier for tests, docs, and bounded fixes.
Optional local fallback (Ollama + Cline or Aider) for outages, quota caps, or a sanitized practice repo.
One AGENTS.md source of truth, branch protections intact, and a shared review checklist before merge.
Tradeoffs
Hybrid routing adds prompt and config maintenance — document which tasks stay cloud vs local.
Dual-running Cursor and Claude Code on the same branch without worktrees creates collision risk.
Startup velocity dies if generated patches skip review; measure accepted patch rate, not prompt count.
A procurement-aware pattern for regulated teams: Copilot or Cursor on the enterprise path, MCP allowlisting, SSO lifecycle, and explicit human review gates tied to the rollout checklist.
Security, platform, and engineering leaders rolling out coding agents under DPA, residency, and audit requirements.
Suggested stack
GitHub Copilot Business/Enterprise or the Cursor plan whose current entitlement includes the required SSO, provisioning, and organization controls; verify each capability against the selected plan.
MCP and extension allowlist: fail closed on unknown servers, review folder and network scope before install, log tool calls.
Claude Code or OpenAI Codex only after the same security review as the editor — each surface can differ on retention and training defaults.
Human review gate on every agent output: plan, tool log, diff, tests, and rollback path documented in the rollout guide.
Tradeoffs
Enterprise seats plus realistic agent usage often exceed the sticker price — budget seat baseline and metered overage separately.
SSO on the IDE plugin does not automatically cover CLI, cloud agent, or API paths; approve each surface.
Governance overhead is real; the payoff is fewer shadow seats and auditable tool permissions.
