Aaron J. Li<sup>1</sup>, Weijia Xu<sup>2</sup>, Chandan Singh<sup>2</sup>, Zixiao Chen<sup>3</sup>, Jeevana Priya Inala<sup>2</sup>, Charlotte Siska<sup>3</sup>,
Acar Tamersoy<sup>3</sup>, Bin Yu<sup>1</sup>, Ion Stoica<sup>1</sup>, Michel Galley<sup>2</sup>, Jianfeng Gao<sup>2</sup>
<sup>1</sup>UC Berkeley <sup>2</sup>Microsoft Research <sup>3</sup>Microsoft Security
Our measurement on SWE-Bench tasks: 18.4–30.9% lower mean cost across five agent–model combinations, cheaper in all 30 repository × combination cells, and Pass@1 up in 4 of 5. Results
TokenEconomizer makes coding agents cheaper to run in the repositories you work in often, without changing the agent itself. It learns from the sessions an agent has already run in a repository, and uses that experience to cut waste in the sessions that follow. It has three mechanisms:
- Micro-skills: remember how the repository works. Sessions keep rediscovering the same
facts: a build step that must run first, the spelling of a command that works here
(
python -m pytestwherepytestis missing,npx tscwheretscis), an environment variable a command needs, where generated files live. Micro-skills learns such facts, in any language, from two sources: failures that repeat across sessions, and a model that reviews the recent sessions everyk sessions (k = 3) and proposes facts. A fact is stated at the start of each session once three sessions have confirmed it, and dropped when later sessions contradict it. - Tool pruning: stop tools that go unused. Every loaded tool's definition is sent with every request. A tool that goes unused overp prompts spread across at leastm sessions, or overs sessions of any length (p = 10,m = 2,s = 5), is left out of the next sessions through the agent's own settings. One command brings it back.
- Failure-loop intervention (the loop guard): step in when a turn is stuck. When a turn (the agent's work on one prompt) keeps hitting the same failures, a judge model checks whether it is stuck. If the judge is at leastc sure (c = 0.7), the agent gets a redirect (a short note, added to its context, to try another approach), at mostr times per turn (r = 1). If the turn is still stuck after that, TokenEconomizer tells you how to stop it; it never stops a turn itself.
TokenEconomizer is a plugin for Claude Code, Codex, the GitHub Copilot CLI and Pi,
with experimental support for opencode, Cursor Agent and Hermes Agent. Each italic
letter above is a setting, shown with its default (Settings), and each mechanism can be turned on or
off. The plugin announces each action in one line, and if any part of it fails, your session
carries on as if it were not installed. Its model calls (the fact proposer, and the judge with the
redirect it writes) run through your own agent, on the same plan or login, and their cost is
included in its reports. What it learns stays on your machine, in ~/.tokeneconomizer.
The paper ran five agent–model combinations on the same 277 SWE-Bench tasks from six repositories.
Each task ran as its own session, and each repository's tasks ran in order, starting with nothing
learned. Costs include every model call TokenEconomizer made. Within each repository and
combination, cost fell by 5–40%. The savings held on Web-Bench (16–21% lower) and on two
security-auditing benchmarks where every task is in a new repository (6–20% lower). The
project page has interactive charts, per-repository numbers
and nine annotated runs, and experiments/ reruns the SWE-Bench and
Web-Bench experiments.
Coding agents are built to finish tasks, not to keep them cheap. Every turn pays again for the agent's instructions, every loaded tool, and the conversation so far. Developers use the same agent on the same repository again and again, so much of that cost repeats: the same facts rediscovered, the same unused tools sent, the same kinds of failing loops. TokenEconomizer leaves the agent's model, weights and code alone and learns across a repository's sessions instead of treating each one on its own, so whatever the plugin spends is paid back over the sessions that follow. Four principles guide the mechanisms:
- Work through what the agent already reads. Facts enter the context at session start, and tools are removed through the agent's own settings, so nothing depends on the agent learning a new tool.
- Count cost as it is billed. Most tokens in a session are cached re-reads, so cost is computed with cache pricing, and it includes the plugin's own model calls.
- Protect task success. Facts form only from repeated evidence, tools are removed only after going unused, and the loop guard leaves alone a turn that is still making progress.
- Keep every action reversible. A fact that later sessions contradict is dropped, removed tools can be given back, and the loop guard only redirects and tells you: stopping a turn is your call.
You need one of these agents, signed in any way it supports, and Python 3.10 or newer on
PATH. The plugin needs no API key of its own and never handles credentials, so any plan or
login works.
Run these commands in the folder where you want to keep the clone. Claude Code, Codex, the Copilot CLI and opencode run the plugin from the clone, so leave it in place. Then start the agent in one of your repositories.
Claude Code
git clone https://github.com/microsoft/TokenEconomizer.git
claude plugin marketplace add "$PWD/TokenEconomizer"
claude plugin install tokeneconomizer@tokeneconomizer
Start claude, trust the folder (Yes, I trust this folder), and run
/tokeneconomizer:status, allowing its script if asked. Once per machine, and again after Claude
Code updates, run /tokeneconomizer:toolprune refresh-catalog.
Codex
git clone https://github.com/microsoft/TokenEconomizer.git
codex plugin marketplace add "$PWD/TokenEconomizer"
codex plugin add tokeneconomizer@tokeneconomizer
Start codex, choose Trust and continue for the folder, then at Hooks need review
choose Trust all and continue. The hooks take effect from your next Codex session, so quit
(/exit), start codex again, and check with $tokeneconomizer:status.
Copilot CLI
git clone https://github.com/microsoft/TokenEconomizer.git
copilot plugin marketplace add "$PWD/TokenEconomizer"
copilot plugin install tokeneconomizer@tokeneconomizer
Start copilot, choose Yes, and remember this folder for future sessions, and run
/tokeneconomizer:status, choosing Yes to run its script. Once per machine, and again after
Copilot updates, run /tokeneconomizer:toolprune refresh-catalog.
Pi
pi install git:github.com/microsoft/TokenEconomizer
Start pi and run /te status.
Cursor Agent (experimental)
python scripts/build_plugin.py --out ~/.cursor/plugins/local/
Reload Cursor, confirm tokeneconomizer appears in Customize, then send a prompt in a repository.
Check its state from a terminal with ~/.tokeneconomizer/bin/te --host cursor status. Cursor's
published hooks provide session, prompt, tool-result and completion events. Its transcript format
and per-session token usage are not a supported public interface, so the adapter records needed
hook events locally and reports session cost as unavailable. Cursor support is macOS and Linux
only: on Windows, Cursor runs hook commands through PowerShell, and this plugin's launcher is a
POSIX shell script.
opencode (experimental)
git clone https://github.com/microsoft/TokenEconomizer.git
mkdir -p ~/.config/opencode/plugins
ln -s "$PWD/TokenEconomizer/plugin/opencode/tokeneconomizer.ts" ~/.config/opencode/plugins/
Start opencode and send a prompt, then check in the repository with
~/.tokeneconomizer/bin/te --host opencode status. opencode has no plugin commands, so the
plugin's commands run from a terminal.
Hermes Agent (experimental; micro-skills and the loop guard, on the classic CLI and one-shot runs)
git clone https://github.com/microsoft/TokenEconomizer.git
ln -s "$PWD/TokenEconomizer/plugin/hermes" ~/.hermes/plugins/tokeneconomizer
hermes plugins enable tokeneconomizer
hermes gateway restart
Link it rather than running hermes plugins install, which leaves the launchers behind when given
this directory and refuses the whole repository at its security scan. Send a prompt, then check in the repository with ~/.tokeneconomizer/bin/te --host hermes status.
If the install worked, the status command prints a short status block. refresh-catalog
records every tool a session loads (without a model call), so that tool pruning can consider all
of them. Keep the marketplace path absolute, as above: the Copilot CLI reads a relative path such
as ./TokenEconomizer as the name of a GitHub repository.
To update:
| Agent | Update |
|---|---|
| Claude Code, Copilot CLI, opencode | git pull in the clone |
| Hermes Agent | git pull in the clone, thenhermes gateway restart |
| Codex | git pull in the clone, then thecodex plugin add line again |
| Pi | pi update --extensions |
plugin/README.md walks through each agent in more detail, including
uninstalling, using Copilot with your own model key, and troubleshooting.
| Agent | Supported | Older versions |
|---|---|---|
| Claude Code | 2.1.300 to 2.1.310 | not supported |
| Codex | 0.153.1 to 0.164.0 | not supported |
| Copilot CLI | 1.0.83 to 1.0.102 | not supported |
| Pi | 1.2.0 to 1.3.0 | 0.99.x: load the extension by path (below) |
| Cursor Agent (experimental) | 3.24.9, macOS and Linux (below) | not tested |
| opencode (experimental) | 1.19.5 and 2.2.0 | not tested |
| Hermes Agent (experimental) | 0.22.5 | not tested |
The commands above were tested on the oldest and newest version in each range; newer versions have not been tested yet.
- Codex before 0.141.0 does not run plugin hooks at all.
- Pi 0.99.x : instead of
pi install, clone the repository and start each session withpi --extension /path/to/TokenEconomizer/plugin/pi/tokeneconomizer.ts, using your clone's full path;git pullin the clone updates it. - opencode support is experimental: it was tested on 1.19.5 and 2.2.0 only, and not in the 2.x terminal UI.
- Cursor Agent support is experimental. Cursor 3.24.9 loaded the local plugin, but hook delivery from a live Agent session has not yet been verified.
- Hermes Agent support is experimental: it was tested on 0.22.5 only, with a scripted local model and one free real model, on the classic CLI and one-shot runs only. Micro-skills and a redirect reaching a real model are verified there; the still-stuck notice reaching a person is not. Its other surfaces (the Ink TUI, the Desktop app, messaging platforms, cron) and subagents are ignored, and tool pruning is not available. Seeits limits .
Work as usual; the plugin needs nothing from you. To see or change what it does, use its commands. They have the same names on every agent; only the way you type them differs:
| Agent | Type a command as |
|---|---|
| Claude Code, Copilot CLI | /tokeneconomizer:<command> |
| Codex | $tokeneconomizer:<command> |
| Pi | /te <command> |
| opencode | from a terminal: ~/.tokeneconomizer/bin/te --host opencode <command> |
| Cursor Agent | from a terminal: ~/.tokeneconomizer/bin/te --host cursor <command> |
| Hermes Agent | from a terminal: ~/.tokeneconomizer/bin/te --host hermes <command> |
| a terminal, in the repository | ~/.tokeneconomizer/bin/te <command> |
| Command | What it does |
|---|---|
status |
what it learned and did here, and what your sessions cost |
microskills |
the facts it injects, and those still forming |
toolprune |
the tools it left out, and the evidence (not on Pi, opencode or Hermes Agent) |
toolprune restore |
give them back; they will not be removed again |
config |
every setting in effect |
config set microskills=on |
turn a mechanism on (or =off ) in this repository; add--global for every repository |
Inside an agent, a command costs one model turn; from a terminal it is free. The other commands
(repos, doctor, config unset, ...) and each agent's permission prompts are listed under
Commands.
Tool pruning and the loop guard are on by default; with these defaults, the plugin calls a model
only when a turn shows signs of looping: the judge, and the redirect it writes if it finds the
turn stuck. Micro-skills is off by default, because its proposer adds model calls and how much it
saves depends on the repository and the work you do there. Each mechanism has its own switch
(microskills, toolprune, loopguard) that takes on or off, and the italic letters above
are these settings:
| Setting | Default | What it sets | |
|---|---|---|---|
| k | microskills_propose_every |
3 | sessions between runs of micro-skills' fact proposer |
| p | toolprune_evidence |
10 | prompts a tool must go unused before it is left out |
| m | toolprune_min_sessions |
2 | sessions those prompts must span |
| s | toolprune_sessions |
5 | or this many sessions, of any length |
| q | toolprune_quiet_minutes |
60 | minutes your sessions here must be idle before a deny list is written |
| e | loopguard.envelope_min_samples |
5 | turns here that ended well (no failing check after the last edit) before the loop guard acts |
| c | loopguard.judge_confidence |
0.7 | how sure the judge must be that a turn is stuck |
| r | loopguard_nudges |
1 | redirects per turn before the guard only tells you |
| b | loopguard.backstop_mult |
2.5 | how many times its usual length a churning turn may run before it is redirected without the judge |
Change them with config set, shown here in Claude Code's form (other agents type it as in
Everyday commands):
/tokeneconomizer:config set microskills=on # turn micro-skills on
/tokeneconomizer:config set microskills_propose_every=5 # k = 5: propose facts less often
/tokeneconomizer:config set toolprune_evidence=20 # p = 20: wait for twenty unused prompts
/tokeneconomizer:config set loopguard_nudges=2 # r = 2: redirect twice before telling you
Tool pruning counts prompts so that long sessions weigh more, and requires m sessions so that a single session never decides alone. It writes the deny list only once your sessions in the repository have been idle for q minutes (or, at the latest, a day after the list formed): a conversation resumed after its tools change must write its whole history to the prompt cache again instead of reading it, and after an idle hour that cache has expired anyway.
On Claude Code, the plugin can also learn from a repository's earlier sessions, read from Claude
Code's own logs oldest first and without any model call. This is off by default, and a
repository's history is read only before the plugin has learned from any session there, so turn
it on right after installing: with config set history_import=on --global, the first session in
each repository reads that repository's earlier sessions in the background (see Earlier
sessions). A deny list that rests on those earlier sessions
waits for toolprune apply, even under auto.
Besides on and off, two modes are useful: toolprune=propose only proposes the deny list, and
nothing changes until you run toolprune apply; microskills=envfacts learns facts only from
repeated failures, with no model calls.
To change a single session, set an environment variable when you start the agent:
TE_DISABLE=1 claude turns the whole plugin off, and TE_LOOPGUARD=off codex turns off one
mechanism (likewise TE_MICROSKILLS and TE_TOOLPRUNE). Every setting and its default are
listed under Settings; each one's environment variable is TE_ plus
its name in capitals.
Micro-skills and the loop guard work the same way on every agent. Each agent learns on its own: facts, tool evidence and the loop guard's sense of a usual turn are kept per agent and repository, while settings are shared. Tool pruning acts through each agent's own settings, so what it can leave out differs:
| Agent | Tool pruning leaves out | Written to |
|---|---|---|
| Claude Code | single tools | .claude/settings.local.json |
| Codex | capability packages | .codex/config.toml |
| Copilot CLI | whole MCP servers | .github/copilot/settings.local.json |
| Pi | nothing (Pi loads only four tools) | — |
| Cursor Agent | nothing yet | — |
| opencode | nothing yet | — |
| Hermes Agent | nothing yet | — |
Tool pruning adds only its own entries and never touches deny rules you wrote; a settings file it
creates is kept out of git (.git/info/exclude). On Claude Code, the session-start line that says
what it left out shows only in the detailed transcript (Ctrl+O), so check with
/tokeneconomizer:toolprune. toolprune restore removes the plugin's entries and nothing else,
and tools you give back are not removed again. If the plugin cannot edit a file safely (a
.codex/config.toml your repository commits, a settings file that is not valid JSON, a symlink
that leads out of the repository), the list stays a proposal and the session start says why;
toolprune apply tries again. docs/PLUGIN.md records how each agent's
hooks and settings were verified to behave, and docs/ARCHITECTURE.md
how the package fits together.
The loop guard first learns how long a turn in this repository usually runs: it waits until e turns here have ended well, on their own and with no failing check after the last edit. After that, a turn that shows signs of looping (the same failures repeating, or errors piling up) goes to the judge. One case is redirected without the judge: a turn that runs past b times the usual length while churning the same edits through the same failures. If a turn is still stuck after its redirects, you see one line that ends with how to stop it:
TokenEconomizer: this turn still looks stuck after a redirect (judge-stuck@0.82, 46 tool calls). Press Esc to interrupt it.
Where that line appears, and how to stop the turn, in the terminal UI and in a headless run
(claude -p, codex exec, copilot -p, pi -p, opencode run, hermes -z):
| Agent | Interactive: where it shows | How to stop | Headless: where it shows | How to stop |
|---|---|---|---|---|
| Claude Code | in Claude's reply | Esc | in the final reply | Ctrl+C |
| Codex | in the conversation | Esc, then /stop |
not shown | Ctrl+C |
| Copilot CLI | in the conversation | Esc twice | on stdout | Ctrl+C |
| Pi | as a notification | Esc | on stderr | Ctrl+C |
| opencode | as a notification | Esc | on stderr | Ctrl+C |
| Hermes Agent | in the model's reply | Ctrl+C | in the final reply | Ctrl+C |
- Claude Code shows a plugin's own lines only in its detailed transcript (Ctrl+O), so the plugin asks Claude to repeat this one in its reply; the model may word it differently or leave it out.
- Codex : Esc interrupts the turn, but a command it started keeps running until
/stop.codex execprints no hook messages. - Pi : Ctrl+C ends a headless run, but a command it started runs on until it finishes.
- opencode 2.x prints the line to the opencode server's error output, not in its terminal UI.
- Hermes Agent : the plugin asks the model to repeat the line, since no hook was found to reach the person's view; a model may word it differently or leave it out.
In a headless run nobody may be watching, so a flagged turn runs until it ends or reaches the
agent's own limits; afterwards, te status counts the redirects and "still stuck" notices.
| tokeneconomizer/microskills | Micro-skill synthesis and injection: facts learned from finished sessions |
| tokeneconomizer/toolprune | Tool pruning: unused tools and Codex capability packages |
| tokeneconomizer/loopguard | Failure-loop intervention: the per-call guard and its per-repository envelope |
| tokeneconomizer/trace | Each agent's session log, read into one shape |
| tokeneconomizer/plugin | The runtime the hooks and the te command run |
| plugin | The installable plugin: manifests, hook files, launchers, slash commands, the Pi extension, the opencode and Hermes plugins |
| experiments | The harness, launchers and task sets behind the paper's results |
| docs | PLUGIN.md (design, verified agent behavior) andARCHITECTURE.md (how the package fits together) |
| tests | Offline tests for the package and the plugin |
The package is standard-library only. Run the tests with python -m pytest tests (add
experiments/tests for the benchmark side), and build a self-contained plugin bundle with
python3 scripts/build_plugin.py --out dist/marketplace.
The code behind the paper's results -- a harness that runs Claude Code, Codex and the Copilot CLI
on SWE-Bench and Web-Bench, the launchers, and the evaluated task sets -- is in
experiments/, whose README gives the setup and one command per
result. It is not part of the installed plugin; the two share only tokeneconomizer/.
@misc{li2026tokeneconomizer,
title = {{TokenEconomizer}: Experience-Driven Black-Box Cost Control for Coding Agents},
author = {Li, Aaron J. and Xu, Weijia and Singh, Chandan and Chen, Zixiao and
Inala, Jeevana Priya and Siska, Charlotte and Tamersoy, Acar and Yu, Bin and
Stoica, Ion and Galley, Michel and Gao, Jianfeng},
year = {2026},
eprint = {2610.22732},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2610.22732}
}
Apache 2.0; see LICENSE.