localharness vs the coding-agent field
make a small local model do senior-dev work on hardware you already own — coding first, multi-role agents next
| localharness | aider | OpenHands | SWE-agent | Claude Code |
|---|
| runs on owned hardware | yes · GX10 | yes · Ollama | yes · 100+ providers | yes · model-agnostic | no · cloud only |
|---|
| per-token cost | $0 (owned) | free local | free local | free local | $ / subscription |
|---|
| edit method | surgical edit + grep | SEARCH/REPLACE diff | file edit + full shell | ACI windowed + lint | surgical edit |
|---|
| test-feedback loop | yes · project tests | yes · lint+test heal | yes · in sandbox | yes · run_tests | yes |
|---|
| tuned for weak models | yes · core focus | partial | general | ACI guardrails | n/a (frontier) |
|---|
| context engineering | verified window · dynamic in/out split · compaction · bail-guard | repo-map · sets ctx | sandbox · condenser | windowed views | large frontier ctx |
|---|
| weak-model rails | forced-commit · circling fast-fail · best-of-N · planning | edit-format coercion | — | ACI guardrails | — |
|---|
| open | in-house | Apache | MIT | MIT (research) | closed |
|---|
| role scope | coding → multi-role | coding (pair) | coding + browser/QA | coding (SWE-bench) | coding + general |
|---|
Where localharness fits. Everyone else needs one of three things we don't: a strong/cloud model (Claude Code), a human in the loop (aider is a pair-programmer), or a heavyweight general agent (OpenHands). SWE-agent is the research ancestor that proved models need their own interface. localharness targets the corner none of them aim at on purpose — senior-dev output from a small model you already own, with the rails and context engineering built for weak models. Coding is the proving ground; the destination is a multi-role agent harness running your own models as proper agents.
localharness is the youngest and narrowest of the five — this compares design intent and fit, not maturity. The others are battle-tested; we're purpose-built for one underserved corner. Peer capabilities per their public docs (aider.chat, OpenHands, SWE-agent / Princeton), June 2026.