TokenForge

Build big things with Claude Code for a fraction of the tokens.

A long Claude Code session re-sends its whole context on every turn. A build that runs for 150 turns at an average of 80k tokens of context reads about 12M input tokens, even though the code it writes is a tiny part of that. tokenforge changes the shape of the work so that context stays small, without lowering code quality.

• Plan once. Your session writes the contracts (shared types and signatures), short library cheat sheets, and a task plan with a check command for every task.
• Build with disposable workers. tforge runs each task in a fresh, minimal claude -p worker. A worker has no hooks, plugins, MCP servers or skills, and only Read, Write and Edit. It sees the contracts, its own files and nothing else, then exits.
• Tests decide when a task is done. The driver runs each task's check itself. A failure goes back to the worker condensed: runtime stack frames are dropped. The last retry escalates to a stronger model. A final project-wide check runs integration workers.
• Measure it. tforge meter shows where the tokens went, per session.

Where it helps, measured

Against clean Claude Code. Every benchmark run uses Claude Opus 5.5 (claude-opus-5-5) on Claude Code 2.1.296, on both sides. This is verified from the model field of every API response in the saved transcripts, not just the --model flag. bench/ runs the same task on the same repo commit in sandboxed sessions, once with clean Claude Code (no plugins, skills, hooks, MCP or CLAUDE.md) and once with tokenforge 0.9.2 at its defaults (build 3af1944f, git 3a44e31f, lean balanced), and scores quality with hidden tests.

Across all 37 tasks (tokenforge 0.9.2, Claude Code 2.1.296), tokenforge used about half the tokens: median −51%, mean −50%, pooled −52%. It used fewer tokens on all 37 tasks, and cost less on all 37. At list prices, which is roughly what usage limits count (most of the token savings are cache reads, at 0.05× the input price; the 1-hour cache writes Claude Code makes cost 2×), the median is −28%. Median tool calls: −29%. Mean quality per task 98.6 against 98.9 (median difference 0): tokenforge scored lower on 8 tasks, by 1 to 3 points except PHP ticketing (−7, a hidden test on seat holds failed in both its runs) and the Vite page (−5, one run lost the regression check).

Built from scratch	Clean Claude Code	tokenforge	Saved	Quality
Vite front page	337k	75k	−78%	100 → 95
Rust TUI (ratatui)	603k	172k	−72%	100 → 100
Go mock REST API	426k	133k	−69%	97.5 → 98.75
PHP ticketing API (SQLite)	1.13M	372k	−67%	100 → 93.13
C# loans API (ASP.NET Core)	1.07M	353k	−67%	98.75 → 100
Bash backup rotation	711k	257k	−64%	95 → 100
Zig JSON toolkit	844k	355k	−58%	98.53 → 95.6
Perl config linter + merger	615k	264k	−57%	100 → 98.5
F# expense splitter (.NET)	765k	330k	−57%	100 → 100
Java URL shortener (JDK HTTP)	797k	352k	−56%	98.5 → 98.5
Python CLI	301k	134k	−55%	100 → 100
Swift cron tool (SwiftPM)	454k	206k	−55%	100 → 100

Median −54% over 24 projects in 23 languages (tokenforge 0.9.2, Claude Code 2.1.296). The agent gets a spec and an empty repo, and hidden black-box tests check the result (CLI, HTTP contract, headless browser, scripted TUI, lune for Luau). A Ruby log analyzer was also run once per side and is left out of these results: tokenforge used 14% more tokens on it (443k vs 390k), quality 96.25 vs 96.83. bench/benchmark.config.json lists the exclusion and the runs are kept.

In an existing codebase	Clean Claude Code	tokenforge	Saved	Quality
TS filters + CSV export	1.24M	545k	−56%	100 → 100
Follow-up session using memory (Rust)	330k	156k	−53%	100 → 98.12
Cross-module debugging (TS)	207k	102k	−51%	100 → 100
Go scheduled notifications	1.09M	551k	−49%	100 → 100
Rust debugging	200k	106k	−47%	90 → 87
Architecture investigation (TS)	487k	264k	−46%	100 → 100
Rust CLI feature	1.25M	706k	−44%	100 → 100
Scheduled transfers (TS, multi-layer)	1.60M	926k	−42%	94 → 100
Go REST endpoint	623k	380k	−39%	100 → 100
Rust refactor	384k	253k	−34%	100 → 100
PR review (Go)	223k	191k	−14%	98.12 → 96.67

Median −46% (tokenforge 0.9.2, Claude Code 2.1.296). The follow-up task runs a second session on the first one's feature, so it measures the memory hint (tforge recall) as well.

Academic work	Clean Claude Code	tokenforge	Saved	Quality
Proofread a LaTeX manuscript (19 planted errors)	266k	169k	−37%	97.77 → 97.77
Answer questions from a 15-page PDF paper	175k	139k	−20%	100 → 100

Two runs per side, tokenforge 0.9.2, Claude Code 2.1.296. Proofreading: 19 errors planted across a 12-page LaTeX manuscript, scored by how many are found (recall) with a small penalty for false alarms. PDF Q&A: 10 questions answered from a 15-page two-column paper, one of them only from a figure. The PDF task hides pdftotext, poppler, Ghostscript and Python PDF libraries on both sides, as on a typical Windows machine; with them installed, both sides simply convert the PDF and the comparison says little.

Each row is the median of 2 runs per side (5 tokenforge and 11 native runs on six of the existing-codebase tasks, from the lean-level campaign and earlier native runs on the same Claude Code), and single runs vary by about ±20%, so read per-task numbers as rough and the overall result as solid. The 0.7.0 and 0.8.0 results this section showed before (median −52%, Claude Code 2.1.293) are in the git history of bench/reports/. bench/reports/ has the raw data, and bench/scripts/charts.py redraws these charts from it. bench/ reruns everything.

Why it works. Every request re-sends Claude Code's tool and skill definitions, and re-reads everything said so far:

• Fixed context: 16.9k tokens per request by default, 9.7k at tokenforge's default lean level (tokenforge 0.7.0, Claude Code 2.1.293). On short tasks this floor is most of the cost.
• Fewer round trips: each call re-reads the whole context. Batching (several files per call, one-call code reads with tread) built whole projects in 3–9 calls where clean Claude Code took 7–22.
• Smaller tool results: they are 80–95% of context growth in existing codebases, and every one is re-read by every later call. Broad dumps fold, long lines are cut, build and test output is compacted.
• Less code: "every stated requirement, nothing extra", written concisely but readably.

The older measurements below are about tforge workers, not the hooks.

Numbers come from tforge meter (Claude Code 2.1.292, Opus main model). "Input-equivalent" weighs each token type by its relative price: cache writes 1.25×, cache reads 0.1×, output 5×.

Where the tokens really go. In one real 11-hour Rust session, the main thread used 1.4M input-equivalent tokens and its 15 subagents used 17.1M (93%):

• 1,253 calls, each re-reading about 80k of context;
• about 570 of their 900 Bash calls were sed, grep or cat, exploring code;
• their total output was only 29k tokens.

Long-lived workers that explore and keep re-reading a growing context are the pattern tokenforge replaces:

• the planner explores once and hands workers exact line ranges;
• every check cycle starts a fresh worker with condensed errors;
• the foreman is a script that costs zero tokens.

Worker overhead. A tokenforge worker carries 2.5k tokens of fixed context per call. A default claude -p with typical plugins carries 26.2k (10×). Shared context is cached across workers. For haiku tasks, thinking is off by default: on a test-writing task that cut output from 10.6k to 4.3k tokens with the same quality.

Small greenfield builds: not worth it. On a browser image editor of about 1–1.6k lines:

	input-equiv	cost	result
one Opus session	174k	$0.78	all checks pass, 35 tests
tokenforge 0.1 (before the thinking fix)	532k	$1.14	all checks pass, 81 tests

A strong model writes a fresh small app in a handful of big batches, so the code's own output is most of the cost and there is little context re-reading to remove. The forge skill tells Claude to work inline in that case.

What's new in 0.9.0

• Planning line, on by default in main sessions: look at the repo first, plan briefly, never draft code in thinking. In an A/B over 36 paired tasks against the same build without it (a 0.8.0 development build of tokenforge (build c9de32de) on Claude Code 2.1.295): list cost −15% per task (95% CI −21% to −10%), thinking −44%, wall-clock −21%, the same hidden tests passed. TFORGE_PLAN=0 turns it off.
• Disk cleanup: tforge gc removes the per-session scratch Claude Code never deletes from /tmp/claude-<uid>/, tokenforge's own leftovers, stale snapshots and used-up handoffs, by rules, with no model call. It runs in the background.
• Instruction files for every agent: sessions, subagents and workers get the AGENTS.md/CLAUDE.md text Claude Code leaves out, compacted.
• Leaner shell output (tmap 0.5.0): tkit run, tkit batch, tkit eval, tkit test --failed, and more commands routed through them.
• Earlier compaction in 1M-context sessions (autoCompactWindow per lean level): replaying 14 days of sessions, main-session input −40%.
• Works with SpecAudit and other skill plugins that bring their own workflow (use balanced).
• Benchmark cost at list prices (the 0.9.2 runs above): TokenForge cost less on all 37 tasks (median −28%), against 34 of 36 for 0.7.0 (median −22%).

All releases: CHANGELOG.md. What comes next: ROADMAP.md.

Install

Inside Claude Code:

/plugin marketplace add chteau/tokenforge
/plugin install tokenforge@tokenforge

Requirements: Claude Code and Node.js 18 or newer. Subscription (OAuth) logins and API keys both work. tmap needs a release binary for your platform, or Rust to build one.

OpenAI Codex

Codex runs the same hooks. Add the marketplace, install tokenforge from it, then open /hooks and trust tokenforge's hooks (Codex skips plugin hooks until you do):

codex plugin marketplace add chteau/tokenforge

What works there: the Bash router (tkit check, tkit test, tview; refusals of interactive ssh and installers), MCP result distillation, the answer cache, prompt routing, the session-start policy, checkpoints and skills. Codex has no Read or Grep tool, so the Read narrowing and code redirect never fire; the 4-minute cap on polling waits is off (Codex takes a rewritten input only with an approval). Context-budget alerts read Claude Code's transcript format and stay silent on Codex. The dashboard and workers are Claude Code only; the proxy also serves Codex with an API key (see API proxy).

opencode

From a clone of this repository (Node.js on PATH):

bin/tforge opencode install            # ~/.config/opencode/plugins/tokenforge.js
bin/tforge opencode install --project  # .opencode/plugins/tokenforge.js, this project only

It writes a one-line plugin that loads tokenforge from the clone, so git pull updates it; tforge opencode remove deletes it. It works with opencode 2 (the { id, setup } plugin, hooks on ctx.tool and ctx.session) and opencode 1 (the v1 factory), from .opencode/plugins/, and runs the same hook scripts: shell commands go through the Bash router (a refusal shows up as the tool's error, a note comes before the output), and the session-start policy and instruction digest are added to the system prompt, the same text on every request so the prompt cache keeps it. Not there yet: MCP distillation, the answer cache, context-budget alerts and checkpoints, which need opencode's message and session events. opencode already reads CLAUDE.md and .claude/skills.

Updates

tokenforge is a normal plugin. It never patches Claude Code or wraps the claude binary, so Claude Code's own auto-update keeps working as usual. It writes to your Claude Code settings in two places only, both reversible: the lean tool level (tforge lean off) and the proxy address (tforge proxy off). To update tokenforge itself automatically, open /plugin, go to Marketplaces, select tokenforge and enable auto-update. Otherwise run /plugin marketplace update tokenforge whenever you like.

To uninstall, first run tforge proxy off (or /tokenforge:proxy off), then /plugin uninstall tokenforge@tokenforge. Without that step your settings keep sending Claude Code's requests to the proxy, which no longer runs, and every request fails; removing ANTHROPIC_BASE_URL from the env of ~/.claude/settings.json fixes it by hand. Nothing else is left behind except .forge/ folders in the projects where you used it.

Disabling the plugin (in /plugin, or enabledPlugins set to false) is not reported to plugins, so the settings tokenforge wrote would keep acting: hidden tools, the shorter auto-compact window, the proxy address. The proxy handles it: while it runs, it checks enabledPlugins at most every 30 seconds, and once tokenforge@… is false it removes what tokenforge added (never your own entries) and remembers it. The first session after you enable tokenforge again puts it back. If the proxy is not running when you disable the plugin (after a reboot, say), nothing sees it, so run tforge lean off and tforge proxy off first.

Use

Nothing to learn: once installed it works on its own, in the terminal and in Claude Desktop.

• Every new session starts with TokenForge: active · lean balanced · replies full · dashboard http://127.0.0.1:7878/, plus a three-line walkthrough for the first three sessions. It is shown to you only, never sent to Claude (TFORGE_BANNER=0 hides it).
• After /clear it says what survived: TokenForge: checkpoint saved (2 min ago, .forge/snapshots/). If there is no fresh handoff, a compact version of the last checkpoint (last two requests, files changed, start of the last reply; at most 900 characters, about 250 tokens per call) is reloaded so Claude continues without re-exploring (TFORGE_CLEAR_RELOAD=0 turns it off).
• Updates: once a day, in the background, it checks GitHub for a newer version and shows TokenForge 0.7.2 is available… with the command to run (TFORGE_UPDATE_CHECK=0 turns it off).
• Papers, notebooks and LaTeX. tkit pdf gives a PDF's text with page markers (opt-in for every Read with TFORGE_DOCREAD_PDF=1: Claude Code's own PDF reading, ~1.3k tokens a page, was cheaper on a table-heavy paper and keeps figures, so it stays the default). Notebooks are read as cells, with long outputs trimmed and images left out. LaTeX builds print only errors with file:line, undefined references and citations, and a box summary. TFORGE_DOCREAD=0 turns the notebook reading off.
• No permission prompts for its own tools. TokenForge's read-only tools (tread, tview, tkit ctx/diff/debug/deps/check/test, tforge recall) are approved automatically, alone or chained with read-only commands like grep and sed -n. Network and remote tools (tkit http/web/ssh) still ask; the multi-file edit tools (tkit edit/patch/fmt) go through only when the session already accepts edits without asking. Without this, every lookup by name asked for permission, so Claude fell back to cat and sed. TFORGE_AUTO_ALLOW=0 turns it off.

/tokenforge:forge build a browser image editor with layers, three filters, undo/redo and PNG export

The skill scaffolds the project with the ecosystem's generator, writes src/contracts.ts, the .forge/cheats/*.md notes and .forge/plan.json, then runs the workers and reports the result. For small jobs (under about 5 files) it tells you to work inline instead, because planning would cost more than it saves.

Other skills:

Skill	What it does
/tokenforge:handoff	Writes .forge/HANDOFF.md (done, decisions, next steps, file map). Run /clear afterwards: the handoff reloads automatically, and the conversation continues at a fraction of the context.
/tokenforge:meter	Token usage of your recent sessions and worker runs.
/tokenforge:terse	Reply style: full (default), lite or off.
/tokenforge:dashboard	Opens the local dashboard and prints its address.

Terse replies (built in, replaces caveman-style plugins)

On by default. At session start, and again after compaction, tokenforge adds one reply-style rule of about 60 tokens. Nothing is added per prompt. Answers lead with the result and skip background nobody asked for. Code, paths, commands, numbers and negations stay exact. Security warnings and irreversible steps stay in full sentences. Files Claude writes keep their normal style.

Measured with Sonnet on three everyday questions (output tokens):

	fixed context per call	EADDRINUSE	rebase vs merge	rm -rf lockfile in prod
no style plugin	0	937	761	599
caveman 3.1.0	+1.2k	408	259	389
tokenforge terse	+0.2k	353	256	182

The fixed context is re-read on every API call, tool calls included, so the smaller rule matters most in long, tool-heavy sessions.

Switch with /tokenforge:terse full|lite|off. lite keeps short full sentences. The choice is saved in ~/.config/tokenforge/config.json. The env var TFORGE_TERSE overrides it. If you also run another reply-style plugin, disable one of them: both rules would load.

Lean tools

Every request carries the definition of every enabled tool and skill. Lean levels hide the ones a coding session rarely needs. Measured fixed context per request, and total tokens over whole tasks against clean Claude Code:

Level	Hides	Per request	Whole tasks
off	nothing	16.9k	—
on	agent-orchestration tools: Workflow, Monitor, Cron*, ScheduleWakeup, RemoteTrigger, PushNotification, SendMessage, ListAgents, TaskStop, DesignSync, ReportFindings, worktrees, NotebookEdit	11.9k	−37%
balanced (default)	also Claude Code's 19 built-in skills, marked user-invocable-only: Claude no longer sees them, but you can still type them as slash commands. Your own and plugin skills, subagents and web tools stay.	9.7k	−47%
max	also the Skill tool, subagents (Task), WebFetch/WebSearch and Claude Code's git instructions	5.8k	−44%
ultra	also Read/Edit/Write: Claude reads and edits files with Bash (sed -n, cat > file, scripts). No image or PDF viewing.	4.4k	−55%

Whole tasks: 6 tasks (architecture, cross-module and Rust debugging, PR review, two features), 3 runs per level against 11 clean runs per task, tokenforge 0.9.2 without the proxy, Claude Code 2.1.296; quality 96.8 to 98.2 at every level against 95.9. max saves no more than balanced here: it hides the Skill tool, subagents and web tools, which these tasks barely use, and one feature task cost more under it; with 3 runs per level that gap may be noise.

max and ultra hide the Skill and subagent tools that skill plugins such as SpecAudit run on: use balanced with them.

Each level also sets when 1M-context sessions compact (autoCompactWindow): at 400k tokens for on and balanced, 300k for max, 200k for ultra, instead of near 1M. Every request re-reads the whole context; replaying 14 days of sessions, 400k cut main-session input 40% and subagent input 17%, at one compaction per ~340 calls. Models with a 200k window are unaffected. A value you set yourself, or pick with /autocompact, wins, and off removes only tokenforge's. Installs from before this get it once, with a notice.

On first start, tokenforge sets balanced once and tells you so in a message shown to you, not to Claude. It writes only its own entries to ~/.claude/settings.json (permissions.deny, skillOverrides and autoCompactWindow), because plugins cannot set permissions themselves. It never re-applies after you choose a level. To skip it, set TFORGE_LEAN_DEFAULT=off (or on, max, ultra) before the first start. To keep tools a level would deny, list them as "leanKeep": ["TaskStop"] in config.json (or TFORGE_LEAN_KEEP); a rule tokenforge already added for a kept tool is lifted at the next start.

These entries are in your user settings, so they apply everywhere: every project, subagents, claude -p scripts and other plugins' skills and agents. A permissions.deny entry wins over any allow rule, including a project's .claude/settings.json and managed settings, so a project cannot bring a hidden tool back; leanKeep (or a lower level) can. When a tool seems missing in some project, tforge lean shows the level and what it hides.

Change the level on the dashboard's Settings page, with /tokenforge:lean <level>, or with tforge lean <level>. The Settings page also picks which built-in skills Claude may still use on its own. tforge lean off removes exactly the entries tokenforge added, never your own. Changes apply to sessions started afterwards.

Dashboard (local web UI)

A local dashboard starts in the background with your first session. Open it with /tokenforge:dashboard, or with tforge ui in a terminal. The address is http://127.0.0.1:7878/, or the next free port.

• Settings (sidebar): health first (how many recent sessions actually loaded tokenforge, which run an older build, which still fire hooks of removed plugins; plugins load only when Claude Code starts, and /clear does not reload them), then the lean level, which built-in skills Claude may use on its own, and terse replies. The overview shows a one-line banner when something needs attention.
• Overview:
  • usage limits, with the time left until each reset;
  • today, 7-day and 30-day totals;
  • a 30-day chart split by token type;
  • the last 48 hours;
  • 5-hour windows over the last week;
  • usage per model.
• Projects: every folder Claude Code ran in, with sessions, calls and subagent share. Open a session to see its context-per-call curve: every point is re-read by the next call, so the area under the curve is what the session cost. Compactions show up as drops. Below the curve, Where the tokens went ranks the session's tool results by estimated cost (size × later calls that re-read it), so you can see which reads were expensive.
• Memory: the project's past sessions as a graph of sessions, files they edited and recurring keywords, with the same search Claude gets through tforge recall. Click a session for its prompts, commits, files and last reply.
• Code graph: folders and files as a force-directed graph, colored by language. Pan, zoom and drag nodes, hover to light up neighbors, switch on call links between files, filter by path or symbol, and click a file for its callers, callees and outline.

Savings in your status line. On first start, if you have no status line, TokenForge sets one that shows an estimate of what it saved today, e.g. TF −45% today (~1.2M saved), and says so once in the startup notice. If you already have a status line it is never replaced; tforge statusline --setup appends the segment to yours (your command keeps running and printing as before). tforge statusline --remove restores exactly what was there; TFORGE_STATUSLINE=0 turns the automatic setup off. The estimate (lib/estimate.mjs) adds up, for sessions shown in the status line: fixed context the lean level removes from every API call (16.9k tokens at off minus your level's measured size, 0 for sessions started before the last lean change), tkit output kept out of context (re-sent on each later call until compaction when it can be tied to the session, otherwise counted once), and prompts the answer cache answered with 0 tokens (one call at the context size before). Percent = saved / (used + saved), in tokens processed (cache reads count fully). Subagents are not counted. The dashboard overview shows the same estimate.

Limits and resets. Exact 5-hour and weekly usage and reset times only exist in the status-line data Claude Code passes to a status-line command, so the same status line records them (tforge statusline --setup if you kept your own). tokenforge edits your Claude Code settings only for lean tools (once on first start, the default level, see Lean tools, and when you change the level: its own deny and skill entries and autoCompactWindow, plus includeGitInstructions at max/ultra) and for the status line above. Without it, the dashboard estimates the 5-hour window from your session timestamps and labels it as an estimate.

Privacy. The server listens on 127.0.0.1 only and rejects other Host headers, which blocks DNS rebinding. It answers GET requests, plus one POST for the Settings page (lean level, skills, terse), accepted only from its own page (same-origin Origin and a JSON body). It loads nothing from the internet. It reads your transcripts incrementally: only new bytes, with a cache in ~/.cache/tokenforge/. From transcripts it keeps and serves numbers only, never prompt, reply or file text. Three exceptions: the session page's "Where the tokens went" table, which shows the command or file path behind each costly tool result (read on demand, never cached); the Memory page, which shows prompt and reply snippets and file names from that project's sessions (the index behind it, in ~/.cache/tokenforge/memory/, holds those snippets); and a project's own .forge/snapshots/, shown on its project page so you can see what /clear will reload; only files listed in that folder can be requested. TFORGE_UI=0 disables the auto-start, and tforge ui --stop stops it.

API proxy (on by default, optional)

Since 0.9.2 the first session after install turns the proxy on: it starts in the background on 127.0.0.1:7879, and only once it answers, ANTHROPIC_BASE_URL=http://127.0.0.1:7879 goes into the env of your Claude Code user settings. Sessions started from then on go through it (the session that set it does not). It is not set when you already route requests elsewhere (your own ANTHROPIC_BASE_URL, Bedrock, Vertex or Foundry), nor when TFORGE_PROXY=0 is set before that first session. Turn it off at any time with tforge proxy off or /tokenforge:proxy off (only the entry tokenforge added is removed), back on with tforge proxy on; tforge proxy --status says where things stand.

Claude Code reads the address once per session, so a stopped proxy would make every request fail. The hooks restart it at session start, at every prompt and after every tool call (a file read and a signal when it runs); if it cannot be restarted, you get a message saying how to start it or turn it off. Before uninstalling tokenforge, run tforge proxy off (see Updates).

What it adds, measured. On 6 benchmark tasks at lean ultra (tokenforge 0.9.2, Claude Code 2.1.296, 3 runs per task), tokenforge saved 53.9% of tokens against clean Claude Code with the proxy and 54.7% without it, at the same quality. So in these sessions (headless, one prompt, 70k to 750k tokens) the proxy adds nothing measurable on top of the hooks. Long interactive sessions, with expired caches and many re-reads, are where its pruning should matter; that is not measured yet. tforge proxy off turns it off.

Without the default, tforge claude runs Claude Code through the proxy for one session, started in the background if needed; tforge proxy runs it in a terminal, --detach in the background, --stop stops it. Works with a subscription login or an API key: the proxy forwards every header as sent. It does three things:

• Logs every request: sizes of the system prompt, tool definitions and tool results, and the usage the API reports (input, cache reads, cache writes, output). tforge proxy report [--days N] sums them, and also shows:

  • Cache breaks. A request that reads much less from cache than the previous request of the same conversation had in context. The report gives a probable cause: the system prompt changed, the tool definitions changed, an earlier message changed, or the cache expired (more than 5 minutes, or 1 hour with a 1h TTL). It also says how many tokens were written again.
  • Folded Reads. How many Edits hit a file whose Read went out folded, how many of those failed because the text was not found, and how many Reads went back into folded bodies. These say whether folding saves more than it costs.

• Compresses tool results the first time they are sent. A whole-file Read above 12k characters (code, not prose or markup) folds long definition bodies, keeping every kept line's number and Read offset=N limit=M in place of each fold; partial reads, which Edit works from, stay whole. Bash and other text output loses colors and progress-bar redraws, repeated lines become a count, and output above 16k characters keeps its head and tail with the full text saved to a file the model can grep. When the cut output is a search result (path:line: lines or a list of paths), it starts with the match count for every file, or every folder, so the cut part still shows where the matches are. Large JSON loses its indentation, and long arrays keep their first, error-looking and last items.

• Prunes stale context. A tool result made obsolete later in the conversation is replaced by a one-line stub saying why: a file read that was read again whole or rewritten later (Read, cat, sed -n, head, tview, including chains like cat a b; sed -n '1,40p' c; every file must be covered), the body of a file written and then read back whole or rewritten, a command run again identically, and a test or build run (npm test, pytest, cargo test, go test, make, tkit test…) superseded by the same run after a change. A file read goes too once that file was written or edited since (Write, Edit, tkit edit, sed -i, tee): from a turn before the last user prompt at once, from the current turn once the model has not touched the file (read, edit, or a command naming it) for 6 calls. Bodies the model wrote go once it has moved on (3 model calls later, or an earlier turn): a Write's content, a heredoc's body (, also after a prefix like ), an Edit's strings, and the heredoc bodies of a script that ran successfully ( replace scripts); the command keeps its shape and the stub says the file holds it on disk, to be read before the next edit ( keeps them until read back or rewritten). Thinking blocks from turns before the last user prompt are dropped too. The newest result is never pruned, a failed later read covers nothing, a failed write or script keeps its body, and an append () never counts as a rewrite. Shell reads (, , , ) are never compressed or capped by the proxy: the model edits from them.

The prompt cache is why only new results change: if earlier content differs from the last request, the API writes it to the cache again at 1.25× instead of reading it at 0.1×. Each compressed result is remembered by its tool_use_id, so every later request sends the same bytes. A result the proxy first passed unchanged (it was started mid-session) is never touched later, and one that Claude Code itself rewrote passes as sent. If a request cannot be parsed, it goes out unchanged. Once an Edit fails with "not found" on a file whose Read was folded, later Reads of that file are sent whole.

Claude Code keeps ANTHROPIC_BASE_URL for the whole session, so (besides the hooks) tforge claude checks every 2 seconds that the proxy is still running and restarts it on the same port if it stopped. An error in one request is logged in ~/.cache/tokenforge/proxy.log and does not stop the proxy.

To estimate the savings on your own past sessions without calling the API, run node bench/scripts/proxy-replay.mjs [--days 30]. It replays the tool results in your transcripts through the same compression and prints only totals. For the recorded benchmark sessions, node bench/scripts/read-replay.mjs shows what code reads through Bash cost and how often the same lines were read again.

Codex (OpenAI Responses API). The proxy also takes POST /v1/responses and compresses large function_call_output items the same way (shell outputs like Bash), with the same memo, sent on to TFORGE_PROXY_OPENAI_UPSTREAM (default https://api.openai.com). Start it with tforge proxy --detach, then in ~/.codex/config.toml:

model_provider = "tforge"

[model_providers.tforge]
name = "OpenAI via tokenforge"
base_url = "http://127.0.0.1:7879/v1"
env_key = "OPENAI_API_KEY"
wire_api = "responses"

Only outputs after the last message (the current turn's tool loop) are compressed, so history already sent stays byte-identical for the prompt cache. Tested with an API key; a ChatGPT-login session goes to a different backend and is untested.

TFORGE_PROXY_COMPRESS=0 logs only. TFORGE_PROXY_FOLD=0 turns off Read folding. TFORGE_PROXY_CAP sets the Bash output cap in characters (0: none). TFORGE_PROXY_MEMO_MB caps the memo of compressed results (default 64 MB; the least recently used go first). TFORGE_PROXY_UPSTREAM sets where requests go (default: an ANTHROPIC_BASE_URL you already had, else https://api.anthropic.com). TFORGE_PROXY_TTL=5m (or ANTHROPIC_CUSTOM_HEADERS='x-tokenforge-cache-ttl: 5m' for one session; the header is not forwarded) replaces Claude Code's 1-hour prompt cache with the 5-minute one: cache writes cost 1.25× the input price instead of 2×, which pays off when requests follow each other closely (scripted runs, tforge run workers). After any pause over 5 minutes the whole context is written again, so it is the default only for claude -p (user agent sdk-cli), where no one pauses; TFORGE_PROXY_TTL=1h keeps Claude Code's 1-hour cache there too.

Privacy. The proxy listens on 127.0.0.1 only and rejects other Host headers. Its request log (~/.cache/tokenforge/proxy/requests-*.jsonl) holds sizes and token counts only. Two files hold content, both readable by you only and dropped after 14 days: memo.jsonl (the compressed results, so later requests can resend them) and out/ (full outputs that were capped).

tmap: code map (built in, replaces CodeGraph-style indexers)

tmap is a small Rust indexer bundled with the plugin. It parses Rust, TypeScript/TSX, JavaScript, Python and Go with tree-sitter and keeps an index in ~/.cache/tokenforge/tmap. Every command refreshes the index first, re-parsing only files whose content changed (size, times and inode are a pre-check, trusted only 2 seconds after a file's last write, as git does; a cache from another tmap version is rebuilt). It respects .gitignore and skips dependency and build folders even without one.

tmap find <words>        ranked definitions: path:start-end  signature
tmap tree [dir|file]     folders, files and their top-level symbols; a file gives its outline
tmap sym|callers|callees <name>
tmap slice <name>        path:start-end, ready for a forge plan's "reads"

Measured on a 1,228-file Rust workspace:

• full index from scratch: 0.9s;
• each later command: about 25 ms;
• one CodeGraph query: 3.4s.

Output for "where is orbit speed handled":

	output
grep -n orbit	~5,100 tokens
codegraph explore	~6,200 tokens
codegraph query	~220 tokens
tmap find orbit speed	~70 tokens, with the right definition and its line range

Repo map at session start (default since 0.9.2, not benchmarked yet). Most tasks begin by listing the repo, and that listing is re-sent with every later request. Session start now loads the tmap tree outline instead (folders, files and their top definitions, two levels deep; file names only when that is too long), capped at 3000 characters (TFORGE_MAP_CHARS, ~750 tokens). Skipped for repos under 8 files, without the tmap binary, or when tmap takes over 3 seconds. TFORGE_MAP=0 turns it off.

Where it pays off: the forge skill uses it to hand workers exact line ranges, and tmap tree is a fast map for you. In A/B runs of normal Sonnet sessions, a session hint did not get the agent to use tmap. Redirecting Grep and Read to it was mixed: two small wins, and one run where the agent worked around it and used more tokens. Both are therefore opt-in:

• TFORGE_MAP=1 adds a one-line hint at session start instead of the map below;
• TFORGE_REDIRECT=1 answers identifier-like Grep calls, Bash greps for one code identifier, and whole-file reads of large source files from the index. Repeating the identical call always goes through.

First use: the launcher downloads the prebuilt binary for your platform from this repo's releases and checks it against the published SHA-256. If no release binary fits, it builds once with cargo (one to two minutes). TMAP_BIN points to your own binary.

tkit: token-saving tools (built into tmap)

tkit <tool> (same as tmap kit <tool>) replaces multi-step shell work with one call whose output is short and shaped for the model. The tools are ported from token-surgeon and use the tmap index wherever it needed CodeGraph. Every tool has --help.

Tool	Instead of	Does
diff [BASE | --pr N]	git diff + reading files	stat, touched symbols with callers, affected tests, whole-function diff; degrades to -U3, then a hunk index
debug [-- test cmd | --trace FILE]	rerun + grep + read frames	failures, top-frame code, code under test, recent diff
analog 'literal'	grep + opening neighbours	where the literal lives, analog files, one import hop
ctx SYM [--refs | --outline]	grep + Read	definition slice, calls, callers
patch (JSON on stdin)	Edit turn + test turn	atomic find/replace edits, then the narrowest tests
distill -q "question" -- cmd	reading a huge log	Haiku reads it; only the answer comes back (lossy: never for code you edit)
check [--fast] [--changed] [-e]	cargo/tsc/go vet/dotnet build/...	build + lint, one line per diagnostic
test [FILTER]	raw test output	summary + failures only
deps ls|where|api|why PKG [SYM]	browsing node_modules, ~/.cargo/registry, ...	a dependency's API or source location
proj, fmt [--check]	reading manifests, formatting by hand	stack facts; format changed files only
edit [-n] [-d] [-t CMD [--keep]]	several Edit calls	atomic multi-file <<< old === new >>> edits; a failing -t test rolls them back
run [--group | --fuzzy] CMD..	raw command output	colours and progress stripped, repeats collapsed, capped with the full output saved; --group: each grep/rg path once; --fuzzy: log lines differing only in numbers merged
batch 'cmd' 'cmd'..	one Bash call per command	independent commands in one call, each output compacted

check, test, deps, proj and fmt detect the stack at the nearest manifest: Rust, Go, TS/JS, C#, Luau, Dart, Python, Java/Kotlin, C/C++, PHP, Ruby, Swift, Elixir, Zig, Scala. Linux, macOS and Windows are supported.

Hooks (automatic):

• Context budget. Every API call re-reads the whole context from the cache, so cache reads grow with calls × context size. The budget is 50k tokens of context per call, or the session's fixed part (system prompt, tools, skills, hook text, measured on its first call) plus 15k if that is larger.
  • The alert is shown to you; notes reach Claude's context, after tool calls, only with TFORGE_WATCH_INJECT=1 (an injected "stop and /clear" note ended unattended tasks halfway).
  • Over the budget, your messages are never held. Your next message gets an alert, and again each time the context doubles, and .forge/HANDOFF.md is written automatically from this session's snapshots (recent requests, every file changed, the last reply; no model call), so /clear at any moment loses nothing: the next session continues from it. A handoff you wrote yourself (/tokenforge:handoff) is never overwritten. The automatic one is moved to .forge/handoffs/ (the newest 10 are kept) once used: when a session that was given it has saved a snapshot of its own work, or after 14 days. TFORGE_AUTO_HANDOFF=0 turns the automatic one off.
• Snapshots. After every reply, the session is saved in chunks to .forge/snapshots/: each chunk holds its requests, changed files, last reply and context size, and the subagents it ran: their type and task, the files they changed, the commands they ran and what they reported (read from their own transcripts, where most of a delegating session's tokens go). A chunk closes after 6 requests or at a compaction, and each session starts its own. Hooks write them from the transcript, not Claude, so they cost no tokens and are current even when no handoff was written. Chunks are kept by use, with no model call: a chunk's score adds up its writing and every later read (Claude reading it, the reload after /clear, the dashboard), each weighted 1/√(hours since + 1), and is lowered when files it changed are gone or a later chunk changed most of the same files. The newest 6 always stay; the others go once the score falls too low (unread: after about 17 days; 4 once its files are gone, 2 once superseded), and at most 50 are kept per project. The dashboard's project page lists them.
• Memory. tforge recall <words> searches the project's past sessions, built from your Claude Code transcripts: prompts, commit messages, files edited and read, and how each session ended. It returns the few matching sessions and files (--session ID gives one session in full). On startup and /clear, Claude gets one line instead of the old snapshot dump: how many earlier sessions exist and the newest one's first prompt, with a pointer to tforge recall. Past work then costs tokens only when the task needs it, not ~2k re-read on every call. The index lives in and refreshes incrementally. restores the old snapshot loading; turns the hint off. In 's two-session task (a follow-up on the previous session's feature), recall used a median of 234k tokens against 263k with injected snapshots and 413k for clean Claude Code, at full quality (two runs each; rough).

TFORGE_KIT_HOOKS=0 turns off the last four at once.

The tforge command

The plugin puts tforge on Claude's PATH. You can also run it yourself with node <plugin dir>/bin/tforge.

tforge init                 example .forge/plan.json
tforge validate             check the plan, print task order and whether workers or inline work is likely cheaper
tforge run [--only a,b] [--force a,b] [-j N] [--dry-run] [--detach] [--proxy]
tforge wait                 wait for a detached run (up to 9 min per call), print its summary
tforge status               task states and spend
tforge prompt <id>          the exact prompt a worker receives
tforge meter [--last N | --all | files...] [--json]
tforge gc [--dry-run] [--json]
tforge caches [--json]      every cache: what it depends on, when it is reused
tforge route "<cmd>"        how the Bash router would rewrite a command (nothing runs)

Plan format

{
  "version": 1,
  "goal": "Browser image editor",
  "context": ["src/contracts.ts", ".forge/cheats/konva.md"],
  "verify": "npx tsc --noEmit && npx vitest run && npx vite build",
  "defaults": { "model": "sonnet", "retries": 2 },
  "tasks": [
    { "id": "history-test", "spec": "Vitest tests for History in contracts.ts ...", "files": ["src/history.test.ts"], "model": "haiku" },
    { "id": "history", "spec": "Implement History from contracts.ts ...", "files": ["src/history.ts"],
      "reads": ["src/history.test.ts"], "deps": ["history-test"], "verify": "npx vitest run src/history.test.ts" }
  ]
}

Field	Meaning
context	Files every worker sees. They go into the system prompt, so after the first worker they come from the prompt cache at about a tenth of the price.
verify	Final project-wide check. If it fails, integration workers fix it, up to retries times.
tasks[].files	Files the task owns. Every file has exactly one owner. Writes outside them are recorded in the ledger.
tasks[].reads	Extra files inlined for this task only.
tasks[].deps	Tasks that must pass first. A failed dependency blocks its dependents.
tasks[].verify	The task's check, one line (join steps with &&). Without one, the task passes as soon as its files exist. tforge validate warns about destructive commands in it (rm -rf /, git push, curl … | sh, sudo).
tasks[].sensitive	true runs this task's check in a temporary git worktree of HEAD holding the current edits (gitignored files such as node_modules are not copied), removed afterwards, even on timeout. Refused outside git.
isolateChecks	true does the same for every task's check and the final verify.
tasks[].model	haiku, sonnet or opus.
defaults	model, integrateModel, tools, maxTurns (12), retries (2), escalate (true), budgetUsd (1.5 per attempt), timeoutMin (20), verifyTimeoutMin (10), inlineMaxChars (60000).

A task reruns only when its definition (or a dependency's) changes. Use --force after changing a contract.

.forge/state.json, .forge/ledger.jsonl (one line per worker attempt: model, turns, tokens, cost, outcome) and the run files are git-ignored automatically. Commit plan.json and the cheat sheets if you like.

Caches

tokenforge keeps three kinds of cache, kept apart; tforge caches lists each one with what it depends on.

• Computation (the tmap index, compacted instruction files, dashboard totals, the proxy's result memo): derived from files. An entry is reused only while what it depends on is unchanged, compared by content where correctness needs it; size and modification time are at most a fast pre-check. Deleting ~/.cache/tokenforge/ costs only time.
• Memory (the project's session index behind tforge recall, the answer cache, snapshots and handoffs): what earlier sessions did or said. It comes back as notes with their source and time, and the answer cache replays an answer only into the same repo, commit, file contents, dependencies, model and terse level.
• Provider prompt cache: Anthropic's, not stored by tokenforge. tokenforge keeps the text it injects identical until its source changes, and the proxy rewrites old messages only when the saving beats the cache rewrite.

Cache files are written whole or not at all (a temporary file, then a rename), so an interrupted write leaves the old file; a torn or unreadable one is ignored and rebuilt.

Safety

Workers run with --permission-mode acceptEdits and only Read, Write and Edit by default. They cannot run commands. Check commands come from your plan and run on your machine, so read a plan before running it, as you would a Makefile. Run builds on a branch or a git worktree. The workers do not load your settings, hooks, CLAUDE.md or MCP servers. Put project rules that workers must follow into a context file.

Each check runs in the project root (or a throwaway worktree, for sensitive tasks and isolateChecks) with the plan's time limit, and at the limit its whole process tree is killed, including commands it started in the background. Timeouts, retries, turns and budgets must be numbers within bounds: a plan with verifyTimeoutMin: 0 is refused instead of running without a limit. A check that writes files in the checkout (outside build and dependency folders) is reported in the log and the ledger (checkWrites): checks should only read.

Environment

Variable	Default
TFORGE_CLAUDE	claude	Claude Code binary used for workers
TFORGE_BUDGET	50000	Context budget per call, in tokens
TFORGE_BUDGET_FLOOR	15000	Room always left above the session's fixed context
TFORGE_WATCH		0 disables the context budget
TFORGE_WATCH_INJECT		1 also puts the budget notes into Claude's context
TFORGE_VIEW_CHARS / TFORGE_VIEW_FOLD / TFORGE_VIEW_LINES	12000 / 8 / 40	tview: dump size that folds, body lines that fold, file lines that fold
TFORGE_VIEW_MAX	BASH_MAX_OUTPUT_LENGTH − 4000 (26000)	tview: output budget; past it the largest code files are outlined
TFORGE_CHECKPOINT		0 stops writing snapshots
TFORGE_SNAPSHOT_PROMPTS / TFORGE_SNAPSHOT_KEEP	6 / 50	Requests per snapshot chunk; most chunks kept per project
TFORGE_HANDOFF_MAX_AGE_H	336	Ignore older handoffs; an older automatic one is moved to .forge/handoffs/
TFORGE_RELOAD_ON	compact,resume	When to reload beyond startup and /clear (0: neither); reloadOn in config.json
TFORGE_STATE_FILES		Comma-separated globs of state files to reload; stateFiles in config.json
TFORGE_SUBAGENT_SKIP		Agent types tokenforge leaves alone (spec-audit:*); subagentSkip in config.json

Benchmark

bench/ holds a reproducible A/B benchmark: clean Claude Code against Claude Code with tokenforge only, on the same model, tasks, repo commits and toolchains. It has 10 tasks in existing codebases (Rust, Go, TypeScript; features, debugging, PR review, architecture tracing, refactoring) and 25 from-scratch projects in 24 languages, each with hidden tests or a hidden answer key and a validated reference solution. Every run is sandboxed: an empty HOME and config dir, no MCP, and a preflight contamination audit. Token counts come from the API usage in every transcript, tokenforge's own overhead and nested calls included.

cd bench
python3 runner/bench.py env build && python3 runner/bench.py doctor
python3 runner/bench.py run --smoke
python3 runner/bench.py run --all

Results go to bench/reports/benchmark-report.md. See bench/README.md for the method, metrics and scoring.

Planning line A/B (shipped as the default; full report). TokenForge with the planning line against TokenForge without it, plus brief as an ablation: 36 tasks, one run per task and arm, the three arms run at the same time, on a 0.8.0 development build of tokenforge (build c9de32de) on Claude Code 2.1.295. List cost −15% per task (geometric mean, 95% CI −21% to −10%; pooled −19%; lower on 27 of 36 tasks), thinking −44%, total tokens −22%, wall-clock −21% (pooled); quality −0.26 points (CI −1.03 to +0.52), hidden tests 1215/1219 in both arms. Nine tasks cost more (up to +26%, architecture), mostly existing-codebase and Q&A tasks with little thinking to cut; four scored lower and two higher, each explained in the report. The line's own 106 input tokens per session are included.

Experiments that did not ship (TokenForge-only variants, one run per task, kept apart from the main results):

Variant	Tasks	Total tokens vs default	Quality	Verdict
TFORGE_PLAN=brief (policy line: plan briefly)	all 36, in the A/B above	+24% vs the look default; −2.7% vs no planning line	4 lower, 2 higher; hidden tests 1214/1219	list cost +13% vs look (pooled); opt-in only
PDFs read as extracted text (TFORGE_DOCREAD_PDF=1)	PDF Q&A without PDF tools, 2 runs	+37% vs clean Claude Code (376k vs 274k median)	unchanged, but extra calls to recover a figure the text dropped	opt-in only
CLAUDE_CODE_EFFORT_LEVEL=medium	C, Kotlin, Java, Swift, PHP, Go feature	−5% (−36%, −12%, +5%, +23%, −9%, +17%)	unchanged	within run-to-run noise; not shipped

Where output tokens go (provider-reported): thinking was 40% of output in the baseline TokenForge runs and 44% in the A/B arm without a planning line; the rest is the code and commands Claude writes and its replies (about 96% and 4% of that, by characters, an estimate). Replies are already short and the code is the work itself, so thinking was the one large lever left: the look line brought it down to 32% of output (thinking −44%), brief only to 37% (−26%), and brief raised first-request thinking.

Development

npm test                        # unit and end-to-end tests with a fake claude binary
cargo test --release --manifest-path native/tmap/Cargo.toml -- --test-threads=1
cargo xwin check --release --tests --target x86_64-pc-windows-msvc --manifest-path native/tmap/Cargo.toml   # Windows build check from Linux
claude --plugin-dir .           # try the plugin locally
claude plugin validate .        # check the manifests

MIT licensed.