Decree

Agent harness generator

Give every agent a harness that knows your team's decisions.

Run one command in your repo. Decree generates a harness for Claude Code, MCP, TypeScript and Python, and a dashboard you host keeps every version and eval run.

npx decree-harness
Self-host the dashboard

self-hosted dashboard

decree.acme.dev/dashboard/harnesses/acme-orders-agent

[ harnesses ]

Acme Orders Agent

acme-orders-agentsynced 2m ago

Version

v7

main@4f2c9a1 · 2m ago

Tools

18

9 read-only · 5 gated

Subagents

2

12 eval cases

Last eval

92%

11/12 passed · 2m ago

Runs

4 most recent

  • 11/12 passed92%v7 · claude-opus-5 · dry-run tools
  • 10/12 passed83%v6 · claude-opus-5 · dry-run tools
  • 12/12 passed100%v5 · claude-opus-5 · dry-run tools
  • 8/12 passed67%v4 · claude-opus-5 · dry-run tools
score by version

npx decree-harness

~/acme-orderszsh
$ npx decree-harness✓ Scanned acme-orders · 214 files✓ Planned Acme Orders Agent · 18 tools✓ Found 41 decisions: 12 live, 27 proposed, 2 superseded✓ Wrote agent/ · claude-code, mcp-server, typescript, python $ npx decree-harness push✓ Pushed acme-orders-agent v7 · decree.acme.dev

01CLI

One command in your repo. Decree scans it and writes the harness.

~/acme-orders1 / 4 · scan
$ npx decree-harness
✓ Scanned acme-orders · 214 files
✓ Planned Acme Orders Agent · 18 tools · asks before 5 risky tools
✓ Found 41 decisions: 12 live, 27 proposed, 2 superseded.
  Confirm with decree-harness decisions confirm

Example output for an illustrative orders and payments service.

02Scoped decisions

Agents get the decisions for the files they touch. The rest stay out of context.

Most teams keep their rules in one AGENTS.md or CLAUDE.md file and load all of it into every session.

A February 2026 study from ETH Zurich tested repository context files like these on real coding tasks. The files did not generally improve task success, and they raised inference cost by over 20% on average. The authors suggest keeping them to minimal requirements.

Decree keeps each decision separate, with globs for the paths it governs. Before an agent edits src/ledger/adjust.ts, it calls get_decisions and gets two decisions: the ledger ADR and the repo-wide rule on money. The other 39 in decree.json stay out of its context.

arXiv 2602.11988 · ETH Zurich · Feb 2026Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents?Gloaguen, Mündler, Müller, Raychev, Vechev
tool call · get_decisionsmcp
// the agent is about to edit a ledger adjustment
get_decisions { "paths": ["src/ledger/adjust.ts"] }

// result (text)
2 decisions govern these paths, most specific first. Follow them, and cite the id when one constrains your change.

[adr-0014-ledger-writes-go-through-ledgerservice] Ledger writes go through LedgerService
Rule: Write to `ledger_entries` only through `LedgerService`, in the same transaction as the change that caused the entry.
Governs: src/payments/**, src/ledger/** | Source: docs/adr/0014-ledger-writes-go-through-ledgerservice.md

[adr-0021-store-money-as-integer-minor-units] Store money as integer minor units
Rule: Store and pass amounts as integer minor units with a currency code. Never use floating point for money.
Governs: ** | Source: docs/adr/0021-store-money-as-integer-minor-units.md

Example output for an illustrative orders and payments service.

03Dashboard

Every version of your harness in one place. Host it yourself, synced from the CLI and CI.

Self-host it →
decree.acme.dev/dashboard/harnesses/acme-orders-agent

[ harnesses ]

Acme Orders Agent

acme-orders-agentsynced 2m ago

Version

v7

main@4f2c9a1 · 2m ago

Tools

18

9 read-only · 5 gated

Subagents

2

12 eval cases

Last eval

92%

11/12 passed · 2m ago

Versions

7 pushed

  • v7Confirm refund rulelatest1 decision live · prompt edited · main@4f2c9a1 · CI · 2m ago
  • v6Supersede decimal money ADR1 decision superseded · main@b81e03d · CLI · 1d ago
  • v5Add cancel_order and events tools+2 tools · main@09ac7f2 · CI · 3d ago
  • v4Initial harness18 tools · 12 evals · main@5d11e8a · CLI · 6d ago

Click the tabs. Example data for an illustrative orders service. A hosted version on trydecree.com is coming soon.

Version history

Each push of a changed decree.json is a new version, with its commit, branch and what changed.

$ decree-harness push

Eval runs

Scores and per-case checks, attached to the version they ran against. Watch a change move the score.

$ decree-harness eval --push

CLI and CI

Approve a login from the browser once, or add an API token to CI and push on every merge.

$ decree-harness login --url <dashboard>

04decree.json

One file in your repo is the source of truth. Reviewed like any other code.

decree.jsontrimmed
{
  "name": "acme-orders-agent",
  "targets": ["typescript", "python", "mcp", "claude-code"],
  "tools": [ /* 18 tools from routes and scripts, plus get_decisions */ ],
  "decisions": [
    {
      "id": "adr-0021-store-money-as-integer-minor-units",
      "title": "Store money as integer minor units",
      "constraint": "Store and pass amounts as integer minor units with a currency code. Never use floating point for money.",
      "status": "live",
      "governs": ["**"],
      "source": "docs/adr/0021-store-money-as-integer-minor-units.md"
    },
    {
      "id": "adr-0009-store-amounts-as-decimal-strings",
      "title": "Store amounts as decimal strings",
      "constraint": "Store amounts as decimal strings with two places.",
      "status": "superseded",
      "governs": ["**"],
      "source": "docs/adr/0009-store-amounts-as-decimal-strings.md",
      "supersededBy": "adr-0021-store-money-as-integer-minor-units"
    }
  ]
}

Every decision has a status.

  • liveServed by get_decisions for paths its governs globs match. Accepted ADRs start here.
  • proposedRules-file bullets and post-mortem lessons. Not served until someone confirms them.
  • supersededReplaced or rejected. Kept for history, never served. supersededBy names the replacement.
~/acme-orderszsh
$ npx decree-harness generate --targets all

decree.json     tools, guardrails, targets, decisions
.decree/        profile.json, manifest.json, schema.json
agent/          
  claude-code/  CLAUDE.md, .claude/ agents and skills, .mcp.json
  mcp-server/   MCP server with get_decisions and decisions.json
  typescript/   standalone agent (Anthropic TypeScript SDK)
  python/       standalone agent (Anthropic Python SDK)
  evals.json    eval cases

Run it on your repo. Then push it to your dashboard.

Needs Node 20.12 or newer. Set ANTHROPIC_API_KEY to have Claude plan the harness, or pass --offline to plan from the scan alone.

npx decree-harness
Self-host the dashboard

Want the cross-repo decision workspace early? Apply as a design partner.