agentic runtime control

make your AI agents
actually work for you

one place to configure, shape, and control how every
AI coding assistant operates — the way you want.

three tools, one mission

agents that operate the way you want

AgentRC Sync

one config, every assistant

Define how you want your AI coding assistants configured once. Sync translates that into native settings for each assistant — Cursor, Windsurf, Claude Code, Copilot, whatever comes next.

Sync also polyfills missing behavior. When an assistant lacks a feature, we patch and normalize so they all behave consistently.

agentrcskillagentrulepermskill → ruleagent → mcp tool
define once · we handle the rest

AgentRC Shape

speak naturally, configure perfectly

Express what you want in plain language. Shape decomposes your intent into the right mix of skills, rules, and deterministic scripts — then writes the configs so your agents do exactly what you mean.

Continually refine your setup by describing changes in natural language. Shape handles the translation to per-assistant settings.

agentrc shape
"When I'm working on tests, I want my agent to always check coverage before committing, suggest missing test cases, and never touch production config files"
analyzing intent...
rules3
→ "check test coverage before any commit"→ "never modify files in /config/prod/*"→ "flag untested code paths in reviews"
skill1
→ test-coverage-analyzer
sub-agent1
→ test-suggestion-agent
slash command1
→ /check-coverage
script1
→ pre-commit-coverage-gate.sh
decomposed into 7 config items · applied to 4 assistants

AgentRC Flow

smart guardrails, not dumb gates

Current tools treat commands as opaque strings — allow everything or interrupt constantly. Flow understands what a command actually does, how risky it is, and whether it matches your intent.

Safe actions proceed. Narrow restrictions apply where you care. No more YOLO mode disasters or death-by-confirmation-dialog.

without flow
$typical assistant
$ rm -rf node_modules
⚠ Allow shell command? [y/n]

$ npm install
⚠ Allow shell command? [y/n]

$ npm run build  
⚠ Allow shell command? [y/n]

$ git status
⚠ Allow shell command? [y/n]
every command = interruption
with flow
agentrc flow
$ rm -rf node_modules
  → deletes: node_modules/
  → risk: low (regenerable)
✔ approved

$ npm install
↻ redirectedpnpm install
✔ approved

$ pnpm run build
✔ approved (build script)

$ rm -rf node_modules/../../..
  → resolves to: /Users/you/
  → deletes: ~93,000 files
✖ blocked path traversal
approve · redirect · block

how it works

from your intent to agent behavior

01you describe what you want
your intent
"I want my agents to be helpful but safe.
They should be able to run builds and tests 
freely, but ask before modifying configs or 
making network requests to external APIs."
02agentrc translates and syncs
generated configs
.cursor/rules.md
.claude/settings.json  
.windsurf/config.yml
.github/copilot-settings.json

→ all expressing the same intent
→ in each tool's native format
03flow enforces at runtime
live session
$ pnpm test           ✔ approved
$ pnpm build          ✔ approved
$ curl api.stripe.com ? external API — allow?
$ vim tsconfig.json   ? config file — allow?
your rules, enforced semantically

ready to take control?

stop fighting your AI assistants.
make them work the way you want.

npm install -g @agentrc/cli