Does AgentWorth upload my code or transcripts to the cloud?

No. AgentWorth is local-first: it scans dotfiles on your machine and stores an index in a local SQLite database, and your data never leaves it. Nothing about you or your work is sent anywhere by default.

Does AgentWorth ever talk to the network?

Two narrow, named exceptions, and neither sends anything about you. `agwt search` can use the optional fastembed feature, which downloads an embedding model from Hugging Face once, the first time you run it -- a dependency fetch, like npm install, off by default in shipped binaries. `agwt blunder --submit` is an explicit opt-in command that posts a single blunder report when you run it. Nothing else reaches the network.

Which AI coding agents are supported?

AgentWorth supports 21 native adapters: Claude Code, Google Antigravity (agy), Cursor Composer, OpenAI Codex, Block Goose, Pi, Herdr, Nous Hermes, OpenClaw, xAI Grok, OpenCode, DeepSeek Code, Kimi Code, MiniMax, Qwen Code, Zhipu/CodeGeeX, Aider, Cline & Roo-Code, Windsurf/Cascade, and Manus.

How do I run AgentWorth?

Run 'npx -y agentworth@latest scan' to index the agent logs already on your machine, or install the native binary with 'curl -fsSL https://agentworth.dev/install.sh | sh'.

Runs on your machine. Sends nothing anywhere.

You’re burning tokens you can’t see.Archie reads the receipts.

Which model made the mistake, what it cost, which run to trust. The answers are in transcripts nobody reads: Claude Code, Codex, Cursor and eighteen others write down everything they do. AgentWorth reads those logs and turns every run into a receipt — tokens, cost, blunders, and what was actually verified.

$npx -y agentworth@latest scan

or curl -fsSL https://agentworth.dev/install.sh | sh

It never phones home. No account, no server, no upload — it reads files that are already on your disk and writes an index next to them.

* * * FLIGHT RECEIPT * * *one laptop · Feb to Sep 2026

SESSIONS INDEXED4,800

HARNESSES ON THIS MACHINE3 of 21

TOTAL TOKENS67.7 B

├─ input419.5 M

├─ output203.0 M

├─ cache writes1.2 B

└─ cache reads65.9 B

AT API LIST PRICE$28,447

VERIFIED: TEST, COMMIT OR CI2,265 (47.2%)

TOP MODELclaude-sonnet-5

[flown] read from local disk, nothing uploaded. Yours prints with archie stats.

Archie reads them.

archie scan indexes every agent history on the machine. archie stats prints the receipt: tokens by kind, cost, the models and harnesses behind them, and how many runs left proof.

> archie statsarchie stats                                       ~/.agentworth/agentworth.db------------------------------------------------------------------------------   4,800 sessions           2026-02-15 -> 2026-09-12           1,827,261 events   EVIDENCE LADDER                                           SESSIONS     SHARE  ----------------------------------------------------------------------------  5  CI or deployment verified ...........................       164      3.4%  4  commit observed .....................................     1,828     38.1%  3  test or build passed ................................       273      5.7%  ---------------------------- the evidence line -----------------------------  2  artifact changed ....................................       345      7.2%  1  done claimed ........................................         4      0.1%  0  unflown .............................................     2,186     45.5%  ----------------------------------------------------------------------------     VERIFIED    rung 3 and up                                 2,265     47.2%   TOKENS                                                           67.7B TOTAL  ----------------------------------------------------------------------------  cache read  #############################################.    65.9B   97.4%  cache write #.............................................      1.2B    1.7%  input       ..............................................    419.5M    0.6%  output      ..............................................    203.0M    0.3%   ADAPTERS                   MODELS                     TOOLS  -------------------------  -------------------------  ----------------------  claude_code    3,866  81%  sonnet-5            1,927  Bash           161,552  codex            745  16%  opus-5              1,008  Edit            26,172  opencode         177   4%  opus-4-8              339  Read            25,506   Next  archie session list --limit 20   the newest sessions, ladder first

The same laptop as the receipt above, on 2026-09-12. Three harnesses, seven months, one command.

> archie session list --limit 20archie session list --limit 20                        4,800 indexed - 20 shown------------------------------------------------------------------------------   EVIDENCE  SESSION         ADAPTER      MODEL        SCORE      DUR    TOKENS  ----------------------------------------------------------------------------  .....     f7009bbe-c2f..  claude_code  fable-5-1       30   1m 38s    194.8K  ####.     agent-a38379..  claude_code  sonnet-5        74   2m 10s      1.2M  ####.     agent-a15d73..  claude_code  sonnet-5        82   2m 19s      1.2M  ####.     agent-a9b94f..  claude_code  sonnet-5        76      53s    395.8K  ####.     1b260767-085..  claude_code  fable-5-1       78   4m 01s    637.2K  .....     ses_f6b7e8bc..  opencode     mimo-v2.5..     23       0s    105.8K  ####.     agent-a62803..  claude_code  sonnet-5        74   2m 09s    826.3K  .....     rollout-2026..  codex        6-astra         26       9s     22.1K  .....     agent-af0651..  claude_code  sonnet-5        23   2m 00s    670.5K  ####.     agent-a81ff7..  claude_code  sonnet-5        74   1m 11s    368.6K  ####.     agent-afbc9e..  claude_code  sonnet-5        67   1m 59s      1.3M  .....     fb73b39e-b68..  claude_code  fable-5-1       27   6m 47s     56.6K  ##...     agent-abeaf5..  claude_code  opus-5          54   7m 10s      2.4M  .....     ses_f6b80fb8..  opencode     mimo-v2.5..     23       0s    106.1K  .....     ses_f6b813ee..  opencode     mimo-v2.5..     23       0s    105.8K  ####.     agent-a264c9..  claude_code  sonnet-5        89   4m 39s      4.0M  .....     ses_f6b8fdef..  opencode     mimo-v2.5..     23       0s    105.5K  .....     ses_f6b8fef4..  opencode     mimo-v2.5..     23       0s    105.5K  .....     ses_f6b902f0..  opencode     mimo-v2.5..     26      15s    213.3K  .....     ses_f6b90479..  opencode     mimo-v2.5..     26      12s    213.9K  ----------------------------------------------------------------------------  8 of these 20 left evidence; the rest are still claims.  archie session show agent-a264c97ee564b423b

Twenty of those sessions, one line each. The dots are the same ladder, one row per run.

Then open it and look.

archie serve runs a local explorer on your own index. Every run, what it cost, what it claimed, and the evidence behind the claim.

The AgentWorth explorer showing one session, its evidence ladder, and the commands behind each rung quoted from the transcript. — One real session: a 25-minute claude-sonnet-5 run on this repo, with the command behind each piece of evidence quoted from its transcript.

21 harnesses

One index across all of them, so the answer to “what has been running on this machine?” is a single command.

Claude Code · Codex · Cursor · Antigravity · Gemini · Goose · Aider · Cline & Roo-Code · Windsurf · OpenCode · Grok · Kimi · Qwen · DeepSeek · MiniMax · Zhipu · Manus · Hermes · OpenClaw · Herdr · Pi

Blame, for agents

archie repo blame src/main.rs names the session, the model and the prompt behind a line of code — months after whoever ran it forgot.

Where the tokens went

Input, output, cache reads and cache writes, rolled up by day, week or month. Your own agent can ask too: claude mcp add agentworth -- agentworth mcp.

Point it at your own machine.

$npx -y agentworth@latest scan

or curl -fsSL https://agentworth.dev/install.sh | sh