summonr

Summon any model.

The open-source coding agent that stays light on your tokens and lighter on your machine.

curl -fsSL https://summonr.dev/install.sh | sh

stats · GPT-6.1 Sol
summonr mid-session on GPT-6.1 Sol: after reading stats.py and running the failing tests, it takes /goal make the tests pass, fixes mean and median, runs pytest to 4 passed, and ends with the goal achieved.
3× fewer tokens than Codex
5–16× less memory
28 providers built in

vs Codex · GPT-6.1 Sol, ChatGPT plan · Terminal-Bench 2.1 · 89 tasks × 3 trialssummonr 0.1.0, Codex 0.162.1

Input tokens per solved task 3× less

summonr
58,976
Codex
182,691

Memory 10× less

summonr
27 MB
Codex
276 MB

Resolution rate*

summonr
85.8%
Codex
79.0%

vs Grok Build · Grok 4.7, SuperGrok plan · Terminal-Bench 2.1 · 89 tasks × 3 trialssummonr 0.1.0, Grok Build 1.0.52

Input tokens per solved task 2× less

summonr
373,831
Grok Build
803,466

Memory 5× less

summonr
27 MB
Grok Build
146 MB

Resolution rate*

summonr
74.2%
Grok Build
70.4%

vs pi, OpenCode · GLM 5.3 Flash, z.ai · Terminal-Bench 2.1 · 86 tasks × 1 trysummonr 0.1.0, pi 1.0.0, OpenCode 2.0.22

Input tokens per solved task 2× less

summonr
95,271
OpenCode
228,578
pi
240,762

Memory 5–16× less

summonr
24 MB
pi
134 MB
OpenCode
394 MB

Resolution rate

summonr
73.3%
pi
72.1%
OpenCode
67.4%

Each pair ran on the same model. OpenCode ran in a later job than summonr and pi. *Solves that looked up published answers count as misses: Codex 11, summonr 0; in the Grok run, summonr 10, Grok Build 5. How we measured

Your agents talk to each other.

Mention any agent with @ and it picks up the message, even on a different model.

builder · Grok 4.7
builder · Grok 4.7
tests · GLM 5.3
tests · GLM 5.3

Three subagents, one build cache.

Each subagent works in its own copy that shares your build cache, and its edits land back in your checkout.

storefront · GPT-6.1 Sol · three subagents
summonr in a terminal on GPT-6.1 Sol: three subagents fix total(), add a --dry-run flag, and document the CLI at once, each line showing what it is doing until it completes, then the tests pass.

Uses the Chrome you already have.

Sign in once in its own profile and it stays signed in. Web search needs no API key.

Browser
The Chrome, Edge, or Brave already on your computer
Profile
Its own, separate from your everyday browsing
Logins
Sign in once in a visible window, kept across sessions
Search
Free, no API key; pages come back as clean markdown

A different model for every job.

Plan on one model, write on another, review on a third.

The summonr workflow card: scout on deepseek-flash, plan on claude-opus-5-5, implement on glm-5.3, review on gpt-6-astra, judge on grok-4.7, summary on glm-5.3-flash.

Your terminal, your theme.

Space, Nord, and Dawn, or your own colors in a theme file. Every pane is drawn by Sigil, our TUI/GUI library.

Theme
A summonr session in the Space theme.The same summonr session in the Nord theme.The same summonr session in the Dawn theme.

Also in the box

Your plan or your key
Sign in with ChatGPT or SuperGrok, or bring a key for any of 28 providers.
Goals
Set a goal and walk away. It stops when your check passes.
Tagged edits
A stale edit fails instead of landing on the wrong line.
No permission prompts
It only asks before damage you can’t undo. There’s no sandbox, so give it a container when it shouldn’t touch your machine.
Add anything
MCP servers, hooks, skills, and plugins, while summonr itself stays at 27 MB.
Private
No account, no telemetry.
curl -fsSL https://summonr.dev/install.sh | sh