Blog · 2026-09-18 · 9 min

Best Local Coding Agent 2026.

Ollama-first harnesses ranked by workflow: Anvaya, Cline, OpenCode, Aider, Continue. Privacy, models, memory, picked by constraint. Updated 2026-09 with Phase 1 measured data.

Direct answer

Local coding agents converged in 2026: Ollama runtimes plus a harness of your choice, zero API spend, code that never leaves your machine. The differences that matter are workflow shape, model appetite, and (the layer most roundups miss) whether anything is remembered between sessions. This edition adds the first measured numbers from our open benchmark program.

How to choose

Start from hardware, not hype. A 7B coder (~5GB) runs on 8GB machines at ~80–85% of 30B quality for everyday edits; 30B-class (~19GB) wants 24GB+ memory and covers refactors; past that you're into workstation territory. Every harness below speaks Ollama's OpenAI-compatible endpoint, so the model decision transfers, the harness decision is about workflow.

Then pick by shape: IDE-dweller (autocomplete + chat in-editor), autonomous agent (plans, edits, runs commands with approvals), or terminal purist (diff-first, git-native, scriptable). No harness wins all three; the honest roundups admit it.

What the measurements now say

Phase 1 of our benchmark program ran six harnesses on the same gateway models across ten task classes with external verification: 238 runs, 228 verified, zero timeouts. Median own-process RSS: anv 25.6 MB (21 MB idle floor), jcode 45.0, codex 97.2, aider 248.9, claude 432.1, opencode 582.3 MB. Node and Python runtimes carry floors no context management removes; Rust harnesses lead on weight.

For the local case specifically, two numbers matter: anv held 4.2 MB kernel footprint at TUI idle with a 12 ms first frame, and measured headless fix tasks cut uncached input tokens 11–21% with 24/24 verified pairs when the memory-aware configuration was active. The full records, including failures, are public.

The field

Anvaya, the memory pick. Local-first Rust agent (14–19 MB static binary, ~5 MB idle, 25.6 MB median harness RSS measured) with a compounding knowledge graph: typed nodes, decay, drift detection, outcome calibration. Fully offline on the Ollama path, 21 tools, subagents, headless CI. Pre-1.0 and building in the open; the one local option whose stated goal is session 50 costing less context than session 1, a thesis Phase 6 measures.

Cline, the autonomous VS Code agent. Approval-gated multi-step loops, MCP support, highest capability and highest token appetite of the IDE bunch. Best for refactors and bug-hunts inside VS Code with a strong local model behind it.

OpenCode, the freedom pick. 75+ providers, TUI plus desktop plus IDE, MIT open source, subagents and plan mode. The default answer to 'I want Claude Code with any model, including local.' No memory engine, sessions restart from summaries. Its Node runtime measured 582.3 MB median in Phase 1.

Aider, the terminal purist. Diff-first edits, auto-commits, architect/editor pipeline, watch mode, famously scriptable. Git history is its memory model; everything else lives in the transcript. Python runtime floor: 248.9 MB median measured.

Continue.dev, the autocomplete-first IDE layer (note: acquired mid-2026; source remains Apache-2.0 to fork). Lowest-friction local Copilot replacement; lightest agent loop of the set.

The missing layer

Every harness above solves privacy, none except Anvaya solves re-learning. A local agent without memory re-pays orientation every session, just privately. That is why the evaluation that matters isn't model-only benchmarks but session-50 behavior: does the tenth session on your repo cost less and know more than the first?

Our position, stated plainly: run any harness you like, and attach memory under it over MCP. Local inference made coding private; compounding memory is what makes it compound. The 2027 refresh will re-rank on measured session-50 numbers, and the memory-benchmark status page tracks exactly what is and is not measured today.

Questions

Asked About This Post.

Which local model should I start with?

Devstral-class default or a Qwen coder at the largest size your hardware runs responsively (7B fallback, 30B where 24GB+ fits). Require tool-call training and 32K+ context; size matters less than harness fit.

Local vs cloud, what do I actually lose?

Frontier edge on the hardest long-horizon refactors, some latency on huge contexts, and managed convenience. You gain zero marginal cost, provable privacy, offline capability, and determinism. Most teams land hybrid: local default, frontier escalation per task.

Are the footprint numbers comparable across tools?

Within the Phase 1 campaign, yes, same models, same tasks, process-tree accounting, external verifiers. Across other sources, no: RSS includes shared mapped pages and hides compressed ones, so we publish both footprint and RSS with the protocol.

Will this ranking change?

Yearly, models, harnesses, and our own measured numbers move fast. This post refreshes with 2027 data; methodology (hardware tiers, workflow shapes, session-50 behavior) stays stable so editions compare.

Run Agents That Fit On Your Laptop.

25.6 MB median RSS. 25 agents ran in parallel on a Core 2 Duo with 4 GB RAM. Hundreds on your machine. Zero cloud required on the Ollama path.

Get Started →Read the FAQ

Requires Rust/cargo to build from source. Linux and macOS today, Windows not yet supported. Pre-1.0, public beta. Pricing TBD.