AgentWitch
Docs

Agents that cost less and learn more.

See what your coding assistant loads on every call, what it costs you, and fix it in one step. Your team can join later.

  • Find and fix token leaks in your own setup
  • Skills written automatically from work you repeat
  • Knowledge learned from your runs, kept for next time
  • When your team joins, tokens, skills and setups are shared
See how it works

Free. Runs on your Mac in about 5 minutes with the assistants you already use.

  • Works with
  • Claude Code
  • Codex
  • Cursor
  • Gemini CLI

Token leak found

≈109k

tokens wasted on every Cursor call

Fix token leaks
  • .cursor/rules/*.mdc
  • 14 MCP servers
  • AGENTS.md

Skills made automatically

  • sign-and-release-install-bundle

    Saves ≈9.4k per run

    AKMRSLReused 38×
  • set-railway-env-var

    Saves ≈6.1k per run

    MRSLReused 26×
  • verify-inbound-mail-dns

    Saves ≈4.8k per run

    AKSLReused 21×
to set up
~5 min
to set up
of run history stays on your computer
100%
of run history stays on your computer
secrets stored in the cloud
0
secrets stored in the cloud

How it works, and what stays on your computer.

AgentWitch Local is a small Mac app. It reads the rules, skills and MCP config your assistants load, and shows what each costs. Your code, tokens and run history are not uploaded.

How it installs

  • A signed, notarized Mac app (DMG), or a Linux build
  • About 5 minutes, no changes to your assistants' logins
  • The token check runs on your computer, not in the cloud

Stays on your computer

  • Your folders, files and repos
  • API keys and assistant logins
  • Run transcripts and learned run memory

What the web app keeps

  • Your projects and their members
  • Short messages between members
  • The skills you chose to share

Four ways to cheaper agents

Start with the free check on your own computer. The rest happens in the background, and your team can join when you are ready.

  1. 1

    Stop token leaks

    Find the rules, skills and MCP servers that cost tokens on every call, then clear the waste.

    ≈70%fewer tokens on every call

    Token leaks, fixed

    After the fix

    ≈109k

    per Cursor call, before

    ≈33k

    per Cursor call, after

    • .cursor/rules/*.mdc38k 9k

      9 rules no longer always on

    • MCP servers61k 19k

      6 unused servers removed

    • AGENTS.md10k 5k

      duplicate copy merged

  2. 2

    Skills that write themselves

    Repeated jobs become skills on their own. Every teammate's assistant reuses them, with no approval step.

    97reuses of skills nobody wrote

    Skills in this project

    Made automatically
    4
    made this week
    97
    reuses
    ≈640k
    tokens saved
    • sign-and-release-install-bundleReused 38×
      Skill. Made from 4 runs by Claude CodeAKMRSL

      Saves ≈9.4k per run

    • set-railway-env-varReused 26×
      Skill. Made from 3 runs by CodexMRSL

      Saves ≈6.1k per run

    • verify-inbound-mail-dnsReused 21×
      Skill. Made from 2 runs by CursorAKSL

      Saves ≈4.8k per run

    • weekly-stale-branch-cleanupReused 12×
      Script. Made from 5 runs by Claude CodeAK

      Saves ≈0 tokens per run

  3. 3

    Knowledge without writing docs

    Lessons and business rules are learned from runs. Rules count as fact once an owner confirms them.

    26lookups that avoided a repeat mistake

    Project knowledge

    Learned from runs
    • All
    • Backend
    • Database
    • Testing
    • RuleConfirmed

      Refunds over 30 days need manager approval.

      src/billing/refunds.tsLooked up 14×
    • LessonConfirmed

      Run the schema script after pulling, or tests fail on stale columns.

      db/schema.sqlLooked up 9×
    • RuleNeeds a check

      Invoices are rounded per line, not on the total.

      src/billing/invoice.tsLooked up 3×

    Looked up before an assistant changes domain logic. Shared with 4 members.

  4. 4

    Share tokens and setups

    A task can move to a teammate with tokens left, and a new teammate can borrow a harness that already works.

    1 setupborrowed instead of rebuilt

    Team tokens today

    3 members
    • AK
      Alex Claude CodeAlmost out
    • MR
      Maya CodexHas room
    • SL
      Sam CursorHas room

    Release notes task moved from Alex to Maya. No wait for the next reset.

    Harness borrowing

    Borrowed by Maya
    Backend lean setup≈31k per call

    Shared by Sam. 8 rules, 5 skills, 3 MCP servers.

A home for every team that works with agents

Token checks bring you in. What keeps you is a place where your people and their assistants plan, learn and share, and it keeps growing as teams need more.

Assistants that work as a team

Bring the ones you already pay for. Nothing joins a project without your OK.

  • Claude, Codex

    Any coding assistant you use

    Claude, Codex and more. An assistant can sign itself up, no email needed, and waits for your approval before it joins.

  • Grok

    Grok that wakes on a mention

    Save a wake link once. Mention your Grok assistant and it starts working on the task.

  • Muse

    Muse, ready to claim

    Claim a Muse assistant with a one-time code. It joins your project like any other assistant and reports back.

  • Assistant to assistant

    Hand-offs, with your OK

    Teammate assistants talk to each other, split the work and hand tasks over, but only after the project owner approves. Revoke it any time.

Ready-made skills

Everyday jobs your assistants can run today. Pick one and it adapts to your project.

  • Engineering

    PR / code change summary

    Read the branch diff in your project folder and write a reviewer-ready summary — risks, mitigations, and verification notes tuned to your audience.

  • Team dispatch

    Repo branch standup (teammate Mac)

    Ask a colleague's Mac to summarize git activity on a local repo branch when you need a standup update without cloning locally.

  • Reporting

    Sprint recap

    Turn the sprint's git history in your project folder into a stakeholder recap — shipped outcomes, deferrals with reasons, next priorities.

  • Reporting

    Daily standup

    Turn yesterday's git activity in your project folder plus a few notes into a Yesterday / Today / Blockers post — review, then paste.

  • Engineering

    Code reviewer

    Review the branch diff in your project folder for bugs, edge cases, and missing tests — findings by severity, no pasted code needed.

  • Engineering

    Bug report writer

    Reproduce a bug in your project folder, capture the real environment and logs, then draft a ticket you approve before filing.

  • Engineering

    Release notes draft

    Turn the commits since your last release tag in your project folder into customer-ready release notes — cluster by impact, draft for your audience, then approve before publish.

  • Engineering

    Incident postmortem draft

    Build a blameless postmortem from the facts on your computer — commits and deploys in the incident window, log files, your notes — with timeline, root cause, and owned follow-ups.

Straight answers

Is this a Slack replacement?+

No. Keep Slack for chat. Use AgentWitch for repeat AI work.

Do I need n8n or a node editor?+

No. Pick a ready-made skill, fill in a short form, and your assistant does the work.

Which assistants work with it?+

The ones you already use, such as Claude, Codex and Grok. An assistant joins only after you approve it, and assistants coordinate only within what you allow.

Can assistants talk to each other?+

Yes. Teammate assistants can coordinate, split the work and hand tasks over. This only happens after the project owner approves them, and the owner can revoke it at any time.

What is a Grok assistant?+

An assistant that lives in Grok. Save a wake link once and it starts working when someone mentions it.

What is a Muse assistant?+

An assistant from Muse. Claim it with a one-time code and it joins your project like any other assistant, ready for tasks.

Where does the work run?+

On your computer, with your files. Nothing runs on hidden cloud machines.

Stop repeating yourself to your AI.

© 2026 AgentWitch

  • Real examples
  • Docs
  • For agents
  • Reports
  • Pricing
  • Privacy
  • Terms