A coding agent harness built for one model: Motif-3.
English · 한국어
Free through September 2026. Motif-3 is served by Infron as Motif: Motif 3 (Free) — $0 per million tokens in and out, the full 262,144-token window — and free access has been announced through the end of September 2026. An account and an API key are all it takes:
npx motifcodeasks for the key, then installs themotifcommand. Terms can change, and the model page is the source of truth.
Not an official Motif project. Motifcode is an independent open-source project. It is not certified, endorsed, sponsored or maintained by Motif Technologies or by Infron. Motif and Motif-3 are their names; this repository is only a client of the model.
Motifcode is a terminal coding agent in the shape of Claude Code — a transcript
above, a bordered prompt below, / commands, @ mentions, a permission prompt,
a .motif/ directory behind it — whose tool set, prompt layout, parser and
failure handling are consequences of things that are specifically true about
Motif-3, most of them
measured rather than assumed. It is not a general harness pointed at a
different base URL.
On 18 August 2026 Korea's Ministry of Science and ICT announced the second-stage result of its Independent AI Foundation Model programme (독자 AI 파운데이션 모델, the national sovereign-model project). Four teams were evaluated; LG AI Research, SK Telecom and Upstage went on to the next round, and Motif Technologies was eliminated — with the highest benchmark score of the four. The vice minister's explanation was that the technology was excellent but that the model "received a somewhat lower evaluation than the other companies on usability and applicability" (사용성·활용성). A week earlier Motif-3 had been released as open weights, scoring 47 on the Artificial Analysis Intelligence Index — first among Korean models. (THE ELEC, Biz Hankook, HelloT; the score has since been revised to 34 following changes to the scoring system)
Usability, for a coding model, is mostly not a property of the weights. It is whether the model's tool calls parse; whether the prompt is laid out the way its own chat template expects; whether one malformed turn is repaired or ends the session; whether the tools it is given are the ones it scored with; and whether the terminal in front of the person behaves like the tools they already use. Every one of those is a property of the harness — and a harness can be written by anyone.
Motifcode was built to prove that the usability an open-weight model was marked down on can be raised far enough with open source. The harness is written for this one model, measured against it, and given an interactive session checked feature by feature against Claude Code's.
|
Wide window. The hero, the welcome card and the prompt. |
 Narrow window. The hero picks the size that fits; the session is the same. |
 First run. The Infron API key is asked for in place of the prompt, checked, and saved once. |
 /skills. The built-in skills, each of them also a command: /commit fix the parser.
|
The hero fits the window; the welcome card says what to type; the first run
asks for the key in place of the prompt. In a session, your line follows >,
the model's prose and each tool call sit behind a ⏺, results under a ⎿,
reasoning stays hidden unless you ask for it, and the status line shows the
context used and how much of the prompt the server served from cache.
Needs Node 20.3+. The package is one file with no runtime dependencies.
cd your-project
npx motifcode # first run: asks for your key, then offers to install the `motif` commandThe first session asks for your Infron API key, once, and saves it to
~/.motif/.env. Because npx leaves no command behind, it then offers to run
npm install -g motifcode for you; say yes and motif (or motifcode) opens
the session from any folder from then on. npm install -g motifcode directly
does the same without the question.
| command | what it does |
|---|---|
motif |
open the interactive session in the current directory |
motif --continue |
the same, with the latest conversation here loaded |
motif "<task>" |
run one task and exit; --interactive stays in the session afterwards |
motif -p "<question>" |
print only the final reply, for scripts and pipes |
motif login · motif logout |
paste a key outside a session; remove the saved key |
motif doctor |
probe the endpoint: auth, tool-call and reasoning parsers, prefix cache, channels |
motif mcp add NAME -- COMMAND [ARGS...] |
register a local stdio MCP server |
motif mcp presets [ID] · install ID |
browse built-in presets and register one offline; --enable opts in to startup |
motif mcp add NAME --transport http URL |
register a Streamable HTTP MCP server; use sse for legacy SSE |
motif mcp list · get NAME · doctor --connect |
inspect MCP configuration and check server connections |
motif mcp enable NAME · disable NAME · remove NAME |
change saved MCP registrations |
motif mcp connect NAME --login · login NAME · logout NAME |
check a connection, sign in through the provider, or clear local OAuth credentials |
motif plugins list · inspect NAME · connect NAME |
discover built-in bundles and review or connect their services; also supports installed skill packages |
motif mcp import codex|claude PATH |
preview server settings from another client; --write NEW_PATH saves disabled entries |
motif sessions · motif resume <file> |
list recorded sessions; resume an interrupted one |
motif skills · agents · plugins · config |
what is loaded, and the effective settings with their sources |
motif skills add · import · marketplace |
install skills from local/Git sources, Claude, Codex and marketplace packages |
motif trust |
approve this repository's .motif/settings.json hooks |
Flags: --model, --endpoint, --env-file, --theme, --thinking,
--verbose, --permissions ask|auto, --cwd, --channel, --max-turns,
--max-output-tokens, --seed, --no-hero. motif --help has the full list.
MCP servers and Skills walk through the mcp, skills
and plugins commands.
If your terminal is configured to render East Asian Ambiguous characters as
two columns, launch with MOTIF_AMBIGUOUS_WIDTH=2 motif to match its cursor and
wrapping behavior. The default is one column. Set this in the shell environment;
it is a terminal preference, not an API configuration value in .env.
MCP servers give Motif outside tools: library documentation, GitHub, a browser,
your own local services. Motif connects to stdio, Streamable HTTP and legacy SSE
servers. Every server sits behind the one mcp tool, so connecting a server
never changes the model's tool list. Each server tool is still called by its own
name against its own JSON Schema, and the arguments are checked before they are
sent. Server tools follow the session's permissions, and a write whose outcome
is unknown is never repeated on its own.
Eight presets ship with Motif. Browsing them is offline, and registering one downloads, starts and signs in to nothing.
| Preset | What it adds | Transport | What it needs |
|---|---|---|---|
context7 |
Current library and framework documentation | HTTP | Nothing; an API key only raises account limits |
github |
Repositories, issues, pull requests and workflows | HTTP | GitHub CLI (gh); Motif reuses the account gh saved |
playwright |
Browser navigation and page inspection | stdio (npx) |
Node/npm and Chrome; runs headless in an isolated profile |
filesystem |
Filesystem tools inside one directory | stdio (npx) |
An existing directory you choose explicitly |
hugging-face |
Public model, dataset and repository information | HTTP | Nothing; a Hugging Face token for authenticated features |
openai-docs |
OpenAI developer documentation | HTTP | Nothing |
tauri |
Tauri app inspection (community server) | stdio (npx) |
A running Tauri 2 app with its MCP bridge plugin |
gmail |
Gmail (preview) | HTTP | Your own Google OAuth access token; no built-in sign-in |
npx presets download their pinned package on the first connection. See
preset requirements for the details.
In a session. Type /mcp. The manager lists the registered servers and the
presets you can add. Select a preset to see what it needs and where it will be
saved, then register and connect it. Its tools are usable in the same session,
with no restart. /mcp list, /mcp connect NAME, /mcp disconnect NAME,
/mcp reconnect NAME and /mcp login NAME work straight from the prompt.
A preset from the shell.
motif mcp presets # the catalog, offline
motif mcp presets playwright # one preset's requirements
motif mcp install context7 --enable
motif mcp install filesystem --root /absolute/project/path --enable
motif mcp install github --enable && motif mcp login githubinstall saves the registration, disabled unless --enable is given.
Any other server. Motif's options go before --, the server's own command
after it. HTTP and SSE servers take a URL:
motif mcp add local -- node /absolute/path/to/server.mjs
motif mcp add local-api --env-ref TOKEN=SERVICE_TOKEN -- node /absolute/path/to/server.mjs
motif mcp add docs --transport http https://developers.openai.com/mcp
motif mcp add remote --transport http --header 'Authorization=Bearer ${SERVICE_TOKEN}' https://example.com/mcp
motif mcp add legacy --transport sse --header-env X-Api-Key=SERVICE_TOKEN https://example.com/ssePass credentials as environment references (--env-ref, --header-env,
${VAR}) rather than values; private headers require them. Registrations are
saved in ~/.motif/mcp.json, which can also be
edited by hand.
From Codex or Claude.
motif mcp import codex ~/.codex/config.toml
motif mcp import claude /path/to/claude-config.json --write ./mcp.imported.jsonImport is a preview until --write names a new file. Imported entries stay
disabled for review, and inline secrets are not copied.
By asking. Paste the server's repository or URL with a clear request, and
the built-in mcp-setup skill takes over:
Connect this MCP to Motifcode and check it: https://github.com/TaewoooPark/Trendchaser-mcp
It reads the server's installation instructions, registers it and checks the
connection under your usual permissions. /mcp-setup <URL> runs it explicitly.
motif mcp list # registrations and presets, offline
motif mcp doctor --connect # start enabled servers, list their tools, close them
motif mcp connect NAME --login # connect, signing in through the browser if needed
motif mcp login NAME --no-browser # print the sign-in URL instead, for SSH or headless machinesServers that use standard MCP OAuth sign in through the browser. Motif opens the
provider's page, receives the callback on a loopback port and keeps the tokens
in private files under ~/.motif/auth, never in mcp.json or in the model's
context. In a session, a server that needs sign-in offers it, and the panel also
shows the URL in case no browser can open. The GitHub preset reuses GitHub CLI's
saved account instead: Motif stores only a delegation grant, gh keeps the
token, and access survives restarts. motif mcp logout NAME removes Motif's
credentials without signing you out of the provider or of gh.
~/.motif/mcp.json is trusted. Project files are never picked up automatically:
review one, then pass --mcp-config FILE together with the SHA-256 it prints to
--trust-mcp. Servers added in /mcp connect at once; edits made from another
terminal need a relaunch. The MCP guide covers tool allowlists,
timeouts, form and URL approvals and the full support matrix.
A skill is a folder with a SKILL.md (YAML frontmatter with a name and a
description, then the instructions) and any scripts, references and assets
beside it. It is the Agent Skills format that Claude Code and Codex use, and
Motif reads both clients' variants. Only a short index sits in the system
prompt; a skill's full body loads when it is used.
Motif ships 18 core skills and 6 more in five workflow bundles. Run one as
/<name> [input] (/commit fix the parser), attach it with @skill:name, or
let the model choose it through the skill tool. /skills lists what is loaded.
| In the source skill | In Motif |
|---|---|
SKILL.md frontmatter, references, scripts, assets |
Loaded with the skill's own directory as its base |
$ARGUMENTS, $ARGUMENTS[N], $N |
Filled from the words after the command; quoted words stay together |
${CLAUDE_SKILL_DIR}, ${CLAUDE_PLUGIN_ROOT} |
The skill's directory, and its package root when installed from a package |
Claude disable-model-invocation, user-invocable |
Respected: user-only, or hidden from the user |
Codex agents/openai.yaml with allow_implicit_invocation: false |
User-only |
Claude allowed-tools |
Kept with a notice; it grants no Motif permission |
.claude-plugin/marketplace.json, .agents/plugins/marketplace.json |
Browsable catalogs for both clients |
| Host-specific hooks, forks, model switches | Reported as unsupported; the skill is not run half-way |
By asking. Paste a skill link with a clear request. The built-in
skill-setup inspects the source, installs the skill and checks it:
Install this skill for this project: https://github.com/anthropics/skills/tree/main/skills/webapp-testing
Asking for it "globally" or "for all projects" installs into ~/.motif/, for
every project. /skill-setup <source or request> runs it explicitly, including
for a Claude or Codex skill you already have.
From a link, a repository or a folder.
motif skills add https://github.com/anthropics/skills/tree/main/skills/webapp-testing
motif skills add https://github.com/anthropics/skills/blob/main/skills/brand-guidelines/SKILL.md --scope project
motif skills add anthropics/skills --path skills/webapp-testing
motif skills add ./my-skill --scope project
motif skills inspect ./my-skill # preview without installingA GitHub folder, SKILL.md or raw link installs the whole skill folder.
--ref pins a branch, tag or commit, --dry-run previews, and the resolved
commit and content digest are recorded with each installation.
From Claude Code or Codex.
motif skills import claude # list candidates; nothing is copied yet
motif skills import codex --json
motif skills import claude --skill webapp-testing --scope projectClaude import reads ~/.claude/skills, the project's .claude/skills and the
plugins enabled in Claude Code. Codex import reads ~/.agents/skills,
~/.codex/skills, the project's .agents/skills up to the Git root and the
plugins enabled in Codex. Skills are copied into Motif's own store; the other
client's files and settings are never changed.
From a marketplace.
motif skills marketplace OWNER/CATALOG
motif skills add OWNER/CATALOG --plugin ENTRY --skill SKILL_NAMEA plugin package with its own MCP servers. motif plugins add OWNER/REPO --skill NAME installs its skills, motif plugins inspect NAME shows the
servers it brings, and motif plugins connect NAME --login registers and
connects them after you approve. Hooks, agents and commands written for another
host stay inactive. /plugin-setup <source or request> does the same by request.
Write your own. Create ~/.motif/skills/<name>/SKILL.md for every project,
or <project>/.motif/skills/<name>/SKILL.md for one; /skill-creator can
draft it.
---
name: explain-widget
description: Explain this project's widget lifecycle and check its invariants.
---
Read references/lifecycle.md relative to this skill's directory.
Explain the widget named in $ARGUMENTS, citing the relevant source.motif skills installed, motif skills update NAME and motif skills remove NAME maintain installed copies. Restart a running session to load newly
installed skills. The skills guide has the precedence rules
and the full compatibility table.
Five plugins ship with Motif. Their skills run as slash commands in every project, and their instructions and references load only when used. A bundle tied to an MCP service is listed for the model only while that server is enabled. User and project skills can override them, and listing plugins never starts a service.
| Bundle | Skills | Optional service |
|---|---|---|
library-docs |
/library-docs |
Context7 |
browser-web-testing |
/browser-testing |
Playwright |
github-workflow |
/github-workflow |
GitHub MCP or an authenticated gh CLI |
frontend-quality |
/frontend-quality, /react-composition |
Context7 and Playwright |
mcp-builder |
/mcp-builder |
The server being developed |
motif plugins list
motif plugins inspect library-docs --json
motif plugins connect library-docsConnecting a bundle's service is a separate step you approve; --yes confirms a
reviewed plan in a noninteractive shell. Existing registrations are preserved,
and a bundle being available does not mean its service is signed in. See
bundled workflows.
Motif-3 is reached over Infron's OpenAI-compatible endpoint. The whole configuration is a base URL, a model id and a key:
| base URL | https://llm.onerouter.pro/v1 |
| model | motif/motif-3 |
| key | MOTIF_API_KEY |
- Sign in at infron.ai/login (email or Google).
- Open Dashboard → API Keys and click Add new key.
- Run
motifand paste the key when asked. It is checked against the endpoint with a one-token request, saved to~/.motif/.env(readable only by you), and never shown to the model.motif logindoes the same outside a session, and/loginand/logoutinside one. AMOTIF_API_KEYin the environment or in a.envnext to the project works too;--env-file <path>puts a file first. motif doctorconfirms the connection and reports how the server returns tool calls and reasoning and whether prefix caching is on.
Free through September 2026. At the time of writing Infron lists the model as Motif: Motif 3 (Free) at $0 per million tokens for input and output, and has announced free access through the end of September 2026. Check the model page and Infron's free-model terms for the current conditions.
Only MOTIF_* keys are read from a .env file, none of them are exported, and
the key is removed from the harness's own environment before anything is
spawned — the agent's bash cannot see it, and neither can a project hook.
| command | what it does |
|---|---|
/help |
commands and keys |
/status (/cost) |
connection, settings and session totals |
/config |
edit settings in the dashboard; /config show lists values and sources |
/stats |
task and tool statistics from local session journals |
/usage |
Infron balance, recorded token usage and reported request costs |
/doctor |
probe the endpoint: auth, parsers, cache, channels |
/mcp [list|connect NAME|disconnect NAME|reconnect NAME|login NAME|logout NAME] |
open the MCP manager to add a preset, or inspect, control and sign in to a connection |
/mcp-setup <URL> |
use the built-in skill to register and check an MCP server |
/login, /logout |
paste an Infron API key, checked and saved to ~/.motif/.env; remove the saved key |
/model [id], /endpoint [url] |
show or set the model id or endpoint for the next task |
/channel [toolcall|object|raw] |
show or set the action channel; changing it restarts the conversation |
/max-turns [n], /max-tokens [n|off], /seed [n|off] |
per-task ceilings and the sampling seed |
/theme [name] |
show, list or set the colour theme |
/thinking |
show or hide the model's reasoning |
/compact [focus] |
replace the transcript with the model's summary of it; words after it say what to keep |
/compact-at [0.5-1] |
the context fraction at which compaction runs (default 0.75) |
/permissions [ask|auto] |
ask before commands, writes and patches run, or run everything |
/cwd [path] |
show or change the working directory |
/notes (/memory), /hooks |
the project notes every task reads; the hooks around tools |
/skills, /agents, /plugins |
what is loaded; each skill also runs as /<skill> [input] |
/skill-setup <source or request> |
inspect, install and verify skills from local files, clients or marketplaces |
/plugin-setup <source or request> |
install a Claude/Codex plugin package's skills and connect its MCP servers |
/new (/clear) |
start a new conversation; the working tree is untouched |
/sessions, /resume [n|file] |
recorded sessions; continue from one |
/quit (/exit, /q) |
leave |
enter send · \ + enter newline · esc interrupt or clear · ctrl-c twice quit · ctrl-d quit
↑ ↓ history · tab show or hide reasoning · ctrl-o output viewer · ctrl-l redraw · shift-tab permissions
@ attach a file · ! run a shell line · # add a project note · / commands · ? hide this
Tool results show a three-row preview by default, including long single-line
JSON and errors. Ctrl+O opens the full output in a read-only transcript viewer.
Use ↑/↓, PageUp/PageDown or Home/End to browse, then Esc, q or Ctrl+O to return
to the same draft and scrollback. This does not truncate the model's tool data.
--verbose or /config verbose true explicitly keeps full inline output.
The Infron balance connects automatically using the key entered at first login.
Press r in Usage to refresh. The account balance and costs recorded in local
journals are separate: unreported or unrecorded charges are not estimated.
| area | what you get |
|---|---|
| Session | Replies stream in as the model writes; reasoning stays out of the transcript unless --thinking or /thinking asks for it; tool calls behind ⏺, results under ⎿; a status line with context used and the prefix-cache ratio |
| Input | @path attaches a file or a directory listing, with a picker as you type; @skill:name attaches a skill's instructions; !command runs a shell line and shows the model its output; #note appends to .motif/NOTES.md; \ + Enter for a newline; long pastes collapsed; Hangul and other wide text handled by display width; ↑↓ history |
| Commands | / opens a menu of every setting; a change made at the prompt is saved to ~/.motif/settings.json; skills run as commands (/commit fix the parser); ? lists the keys |
| Permissions | A numbered prompt before a command, a write, a patch or the terminal runs; "don't ask again for this tool"; a refusal the model is told about; Shift-Tab or /permissions auto runs everything |
| Conversation | Each task sees the ones before it; --continue and /resume bring a recorded conversation back; messages sent while a task runs are queued; Esc interrupts; Codex-style compaction past compactAt of the window — the model writes a handoff summary and your own messages are kept verbatim — and /compact <focus> on demand |
| Backend | .motif/ laid out like Claude Code's .claude/: user and project settings, skills, agents, plugins, notes, one journal per task, history; project hooks applied once motif trust approves them |
| Skills and agents | 18 core skills (explore, plan, explain, code-review, security-review, test-fix, debug, refactor, commit, pr-body, docs, init, skill-creator, skill-setup, plugin-setup, mcp-setup, motif-endpoint, korean); 6 additional skills in five built-in workflow bundles; Claude Code and Codex skills and plugins installed from links, repositories, marketplaces or the clients themselves (Skills); 5 built-in subagents (explorer, reviewer, tester, planner, patcher) |
| MCP | stdio, Streamable HTTP and legacy SSE servers; 8 built-in presets set up from /mcp without a restart; CLI registration and Codex/Claude config import; mcp-setup for setup by request; browser OAuth (--no-browser for SSH), GitHub through gh, human form/URL approval; server tools follow session permissions (MCP servers) |
| Endpoint | The key asked for once and saved to ~/.motif/.env, withheld from every command the agent runs; a 401 that says which side of the key it is on; a 429 retried after the server's Retry-After; motif doctor reports what the server actually returns |
| Screen | Shrinking the window mid-session leaves no stale rows; five themes (motif, claude, mono, solarized, dracula) swapped in place |
| Scripts | motif -p "question" prints only the reply; motif "task" runs one task and exits |
Skills are covered in Skills.
A subagent is Markdown with frontmatter — name, description, tools (a
count, or a prefix of the canonical list), readOnly, maxTurns. A plugin is
a directory with plugin.json and its own skills/ and agents/.
~/.motif/settings.json your defaults: model, endpoint, channel, budgets, theme, thinking, compactAt, permissions
~/.motif/.env the credential
~/.motif/mcp.json MCP server registrations
~/.motif/auth/ MCP sign-in credentials and delegation grants, private to you
~/.motif/skills/<n>/SKILL.md, ~/.motif/agents/<n>.md yours, on every project
~/.motif/skills-installed.json, ~/.motif/skill-packages/ managed skill copies and sources
<repo>/.motif/settings.json the project's settings and hooks — applied once `motif trust` approves it
<repo>/.motif/skills/, agents/, NOTES.md the project's
~/.motif/plugins/<n>/, <repo>/.motif/plugins/<n>/ plugin.json + skills/ + agents/
<repo>/.motif/sessions/*.jsonl one journal per task
<repo>/.motif/history.jsonl what you typed, for ↑
A generic harness assumes the model's tool calls parse, that the tool list is free to change, and that reasoning is optional. None of that holds for Motif-3, and each fact below, checked rather than assumed, became a design constraint.
| Fact about Motif-3 | Source | What it forces here |
|---|---|---|
| The chat template renders the tools block before the system prompt, in the same turn; reordering two tools leaves ~24% of the prefix | chat_template.jinja; measured in template.test.ts |
Nine tools in a frozen, canonical order (done, bash, read, write, apply_patch, term, skill, task, mcp); a subagent takes a prefix of it; a prompt under 3k tokens (about 2k for a one-shot run without the MCP tool) of which 90–98% is served from the endpoint's cache |
Intermediate reasoning is rendered only when tools are registered, and the hosted router does render returned reasoning_content |
template, measured; endpoint, 2026-09-20 | Tools are registered on every channel, and the model's reasoning is sent back every turn |
JSON inside <tool_call> is frequently malformed (shell \$, regex \s), and the hosted endpoint sometimes emits a bare call with no tags |
the vendor's vLLM parser; endpoint, measured | A client-side repair ladder behind the server's own, recovery of bare calls, a breakage budget, and closed tool schemas enforced by a linter that fails the build |
| A dropped tool call and a final answer look the same | vendor parser comments; measured in the campaign | done is a tool; in benchmark mode a turn without an action is handed back instead of ending the task |
SWE-bench Verified 76.2 came from a single bash tool; Terminal-Bench 2.1 74.9 from a persistent tmux session |
mini-SWE-agent config; Terminus 2 | The thin tool set is the baseline, with a term tool beside bash |
| A reasoning step takes 200–300 s on the hosted endpoint | measured | No 300-second header timeout in the transport; a 429 is retried after Retry-After |
The hosted endpoint has no /v1/completions |
motif doctor |
Only the native toolcall channel runs there; the object and raw channels and the pruning toolkit need a local completions server (--experimental-channel) |
Whatever the channel, the model's own body is kept verbatim in the transcript, and the loop is tested by injecting the faults the model actually produces: invalid escapes, truncated calls, unparseable bodies, empty turns, a dead server.
Motif-3 is a 314B-parameter mixture-of-experts model with 13.2B activated per
token and a native 256K context, released under MIT by
Motif Technologies:
weights,
technical report,
serving fork with the motif
tool-call parser, and the Infron API
used here, which also serves /v1/responses and Anthropic-style /v1/messages,
so Codex, Claude Code and other harnesses reach the same model.
Aider polyglot alone cannot settle the comparison. It measures spec-to-code with a test loop in a two-file repository, one run per instance. SWE-bench Verified and Terminal-Bench 2.1, the benchmarks behind Motif-3's published scores, still need to be run with the same three harnesses before the gap can be called general.
The claim above was measured on 2026-09-20/21: the same model behind three harnesses on the Aider polyglot benchmark — 213 Exercism exercises across C++, Go, Java, JavaScript, Python and Rust (225 minus 12 excluded before the run), one attempt per instance, seed 0, a 15-minute cap, the same task text, and the same grader applying each patch to a pristine checkout. Motifcode ran in benchmark mode as shipped; Codex CLI and OpenCode ran through thin adapters against the same Infron endpoint.
| harness | passed | pass rate (95% CI) | vs. Motifcode, paired | McNemar p |
|---|---|---|---|---|
| Motifcode 0.3.0 + Motif-3 | 196 / 213 | 92.0% (87.6–95.0) | — | — |
| OpenCode 1.17.9 + Motif-3 | 177 / 213 | 83.1% (77.5–87.5) | −8.9 pp (−14.6, −3.8) | 0.003 |
| Codex CLI 0.154.0 + Motif-3 | 170 / 213 | 79.8% (73.9–84.7) | −12.2 pp (−17.4, −7.0) | < 0.001 |
The gap is the harness, not the model. Codex ended 38 rows on the time cap and
2 on a malformed tool call echoed back into history; OpenCode ended 21 on the
router's repetition abort and 18 on the cap; Motifcode ended 22 on the cap and
9 on its own turn and loop guards, and 182 normally. One seed at temperature
1.0, so differences under ~8 pp are not resolved; Codex ran with its stream
idle timeout raised (stock 300 s dropped the stream on this endpoint's long
reasoning steps), OpenCode with web tools denied. Method, per-language results,
every row and the incident log are in
packages/eval/REPORT.md; the adapters and scripts
to repeat the run with your own key are in
packages/eval/polyglot-bench/.
packages/protocol/ chat template · tool-call repair · reasoning scrubber · channels
packages/tools/ the frozen tool set and its linter
packages/core/ agent loop · endpoint config · compaction · breakage budget · loop guard
packages/replay/ record, replay and deliberately break the transport
packages/tui/ typed cells · two-region streaming · composer · menus · themes
packages/mcp/ MCP client and manager · presets · OAuth and GitHub CLI sign-in · schema checks
packages/skills/ Claude/Codex-compatible skill parsing, the registry and the built-in skills
packages/agents/ subagent definitions and the local scheduler
packages/hooks/ lifecycle shell hooks
packages/journal/ append-only session log, resume, trajectory export
packages/cli/ the `motif` command, the interactive session, login, doctor, MCP/skill/plugin setup
packages/eval/ polyglot suite, campaign runner, worktree grader, REPORT.md and the polyglot-bench/ kit (the benchmark above)
toolkit/ prompt goldens (jinja2), expert-pruning surgery, campaign score table
corpus/ vendored template + generated goldens
docs/ logo, screenshots, benchmark figure, model guide, MCP and skills guides
pnpm install && pnpm typecheck && pnpm build
pnpm test # unit, integration, CLI end-to-end, and an install smoke
pnpm lint:tools # schema linter — fails the build on loose schemasTo release, bump version in both package.json files and VERSION in
packages/cli/src/main.ts, commit, and push a tag (git tag v0.3.0 && git push origin v0.3.0); the release workflow runs the suite and publishes to npm through
trusted publishing, with provenance and no stored token.
Prior art read closely: Claude Code for the session's shape; Codex for typed history cells, two-region streaming and the compaction handoff; gemini-cli for approval queues and loop detection; hermes-agent for the streaming think-scrubber; Terminus 2 for the persistent-terminal contract; mini-SWE-agent for proof that a thin tool set is enough.
Taewoo Park — physics and spintronics at KAIST, building harnesses for science and code. taewoopark.com · GitHub · X · LinkedIn
Apache-2.0. The Motif-3 weights and chat template are MIT, from Motif-Technologies/Motif-3; any derived checkpoint inherits that licence and credits the original. Motifcode is not affiliated with Motif Technologies or Infron.