Contract
Before work starts: the goal, the paths and effects allowed, the budget and one acceptance test. Six controls apply to any run: tools, sandbox, budget, approval, effects and audit.
Set it up once. Every agent follows it.
Four steps put every agent and chat on the same rules. Each command is in the quickstart.
- InstallAdd the free, open-source CLI.
- Wire a repositoryOne step adds the commit hook and the checks for every file type you use.
- Connect agentsPoint Claude Code, Codex, Cursor or any MCP client at Headron.
- Share with chatGive ChatGPT, Claude, Gemini or Grok one link. It works to the same rules.
What you get
- 9×Fewer tokens
- Each task loads only the rules, files and past failures it needs: 89% fewer prompt tokens.
- 700Mistake kinds refused
- Deterministic checks, not a model, decide pass or fail. One verdict on laptop and CI.
- 353Files mapped to rules
- Each file is tied to the rule it answers for, so a change knows which checks it owes.
- 122Failure ledger
- Every failure shape is logged with its fix. 99 already have a guard, so a repeat is refused.
- 6Controls on every run
- Narrow tools, a sandbox, a budget, approval, side-effect limits and an audit trail.
- 13Rules you write
- Pick a task profile, then set the paths, commands, network and side effects it allows.
- 388Declared tools
- Every tool is declared and validated. A task runs only the commands it lists.
- 94Honest instruments
- Each states what it proves and what it does not. Results carry a receipt tied to the commit.
Measured by the instruments in the README. They show what was caught, not that a task succeeded.
Connect
- MCP
- Connect to public docs or use the local server for project-specific checks. MCP setup
- Any chat
- ChatGPT, Claude, Gemini or Grok follows the rules from one link and says what it has not proved.
- CLI
- Every commit waits for its checks, whoever wrote it. Free and open source.
- CI
- The same checks run on every pull request. Drift fails it.
- Mind API
- Rates each change by risk, lists its checks and routes it to an agent or a person.
- Your agents
- Command line agents, ACP agents and hosted providers, with your own keys and models.
Connect with MCP
Connect an MCP client using Streamable HTTP at https://headron.tech/mcp. The public server offers the docs and agent guide as resources and through read_document. No account or API key is required.
In your client, add a remote MCP server and paste the endpoint above. Ask it to list resources or read the agent guide. The server shares public documentation; it cannot access your files or run checks.
For project-specific routes, checks and failure history, install the local, read-only stdio server instead.
For machine-readable discovery, see Agent guide. The public MCP endpoint does not provide the hosted Mind API.
Mind API
Decides what happens to each change before an agent acts. Agents call it over MCP; people call it from the command line.
- Classify
- Rates a change from low to critical. A change it cannot place is refused, never guessed.
- Plan
- Lists the checks owed, whether a sandbox is needed and how many retries it gets.
- Route
- Routine work to an agent, security changes to a read-only reviewer, critical ones to a person.
- Policy
- Sets required checks, forbidden patterns, a time limit and a context budget.
- Recover
- Retries within budget, rolls back, isolates or escalates with the evidence attached.
- Refuse
- Stops unsafe work with a reason, the rule that blocked it and a fix.
- Direct
- One call returns rating, plan, route, policy and related files. Four round trips become one.
- Search
- Finds related files from a local index. Nothing leaves your machine.
- Receipts
- Lists recent receipts and whether each still checks out.
Free on your machine in the open core. The hosted Mind API, in development, answers the same calls from any computer with one API key and saves handoffs to your account.
Memory
Every refusal goes into a failure ledger. Before an edit, an agent can ask what broke in that file and what fixed it.
122 failure patterns are on record and 99 have a standing guard, so a repeat is refused, not rediscovered.
Verdicts are saved by id, never by paraphrase. An id can be re-checked; a summary drifts.
How a change lands
Checks
Each file type owes checks from format to runtime. A git hook runs them at commit, whoever made the change. Your machine and CI give the same verdict.
Verdicts
- PassRan and met its test.
- FailRan and missed. Blocks the merge.
- SkipNot run or tool missing. Not a pass.
- Not runNever reached. Not a pass.
Receipts
One record per run: the policy, the input and each check with its verdict, bound to the starting commit. Events are chained, so an edit breaks every later one.
Self-checks
- Caught by name
- Each planted mistake must be refused by the check built for it.
- Plans, then code
- The model plans a change; fixed code applies it.
- Stated limits
- Every check says what it proves and what it does not.
- Only tighter
- Quality limits move one way. Loosening one is done in the open.
Limits
Headron proves which checks ran and what they found, for recorded events only. A check is only as good as its test, so keep reviewing what matters.
Questions
Does Headron run my agent?
No. Your agents run as they do today. Headron decides which checks a change owes, holds the commit until they pass and records the verdict.
Which agents does it work with?
Any agent or runtime: Claude Code, Codex, Cursor, Antigravity, OpenHands, Windsurf, Hermes or a plain terminal. Each keeps its own tools.
What happens when a check tool is missing?
The check is reported as skipped and kept out of the pass count. A missing tool can never turn into a pass.
Does it slow agents down?
Routing hands an agent only the checks and lessons for the file in front of it, so it reads less, not more.
Is it open source?
The core is free and open source under the MIT license, including the Mind API's decision logic. Hosting and the paid plans are in development.