Delimit — Keep the state. Change the model.

</> Delimit

Keep the state. Change the model.

Delimit keeps your decisions, tasks, and handoffs on your machine, so the next AI coding session picks up where the last one stopped, even in a different assistant.

Glama directory grades: License A · Quality A · Maintenance A. Observed September 15, 2026, for this public repository. View the listing.

The problem

• Your agent forgets what you decided last session, then asks again or quietly decides differently.
• It redoes work that was already finished, or says "done" when nothing shipped.
• It takes actions you never approved.

Quick start

npx delimit-cli setup

setup adds Delimit to the assistants it finds on this machine. Open your assistant as usual and it can read and write the same saved record. You can also save and find a decision from the terminal:

The same session as text:

$ npx delimit-cli remember "Auth uses session cookies, not JWT. Decided after the token-leak review."

Remembered. (1 memory total)

$ npx delimit-cli recall auth

Delimit Memories

[mem-9586382bff96] just now [cli] ✓ Auth uses session cookies, not JWT. Decided after the token-leak review. #jwt

1 memory found (1 total)

Works with: Claude Code, Codex, Cursor, and Antigravity (configured by setup), plus GitHub Copilot and Muse as full sessions through delimit chat. Install and sign in to each client yourself; see Lead chat, Copilot CLI and Muse Code sessions.

npx delimit-cli chat --model copilot

npx delimit-cli chat --model muse

Hard calls: npx delimit-cli deliberate " " puts a question to the models you have configured (Claude, GPT via Codex, Gemini via Antigravity, Grok, DeepSeek, Muse) and shows you where they agree and where they disagree.

If Delimit helps you, a ⭐ helps other people find it.

For code repositories: merge gate

Delimit can also gate AI-assisted changes in a repository and keep a local signed attestation of each run.

$ delimit wrap -- claude "fix the flaky test in tests/api.spec.ts"

✓ repo_diagnose ✓ security_audit 0 critical · 0 secrets ✓ test_smoke 165/165 ✓ changed_files 1 ✓ attestation signed att_a05050eb8e13277e delimit.attestation.v1 · HMAC-SHA256 receipt → .delimit/attestations/att_a05050eb8e13277e.json

Every wrapped run emits a local delimit.attestation.v1 bundle: repo head before/after, changed files, gate results, and an HMAC-SHA256 signature. Verify it offline with delimit seal-verify; hosted replay is not available yet. Advisory by default; flip to enforcing when you're ready.

---

Zero-config PR check: delimit check

Zero-config PR safety gate. No init, no setup, no account, no keys — run it on any repo and it applies deterministic checks (breaking API changes + leaked secrets) to your staged or modified files.

npx delimit-cli check

$ delimit check

Delimit Check

Policy: default | Base: HEAD | Specs: 1

• api/openapi.yaml — clean

PASSED — no breaking changes, no leaked secrets

When a check finds a breaking change or a leaked secret, it prints the offending file and the issue and exits non-zero — so it drops straight into a pre-commit hook or CI step.

Common options:

delimit check --staged          # only check staged files

delimit check --base main # compare against a git ref (default HEAD) delimit check --fix # show migration guidance for violations delimit check --record # write a content-pinned record of this check

--record writes a content-pinned record of the check to .delimit/records/check- .json (or a path you name). That record is the precursor to the signed, replayable Seal attestation — the same evidence shape, pinned to the exact content you checked.

Methodology · cal.com v2 worked example · Workflow guide · Website

---

Worked examples

Real, reproducible merge-gate runs against public API specs:

• [OpenAI OpenAPI: a year of AI frontier evolution under a cross-vendor merge gate](https://delimit.ai/reports/openai-openapi-attestation) — OpenAI (openai/openai-openapi)
• [Stripe v1 OpenAPI: 57 days under a merge gate](https://delimit.ai/reports/stripe-openapi-attestation) — Stripe (stripe/openapi)
• [Anthropic API: 76 days under a cross-vendor merge gate](https://delimit.ai/reports/anthropic-api-attestation) — Anthropic (anthropics/anthropic-sdk-python)
• [Twilio v2010 OpenAPI: 55 days under a merge gate](https://delimit.ai/reports/twilio-api-attestation) — Twilio (twilio/twilio-oai)
• [Docusign eSignature v2.1 OpenAPI: 46 days under a merge gate](https://delimit.ai/reports/docusign-esign-attestation) — Docusign (docusign/OpenAPI-Specifications)
• [Supabase Auth OpenAPI: 57 days under a merge gate](https://delimit.ai/reports/supabase-auth-openapi-attestation) — Supabase Auth (supabase/auth)
• [cal.com v2 OpenAPI: 60 days under a merge gate](https://delimit.ai/reports/cal-com-v2-attestation) — cal.com (calcom/cal.com)
• [EU TED v3 procurement API: $ref'd component-schema drift under a merge gate](https://delimit.ai/reports/eu-ted-v3-attestation) — European Commission (TED v3 Public API)
• [Cross-agent handoff: one artifact, four CLIs](https://delimit.ai/reports/cross-agent-handoff) — Cross-CLI session handoff (worked example)
• [delimit-mcp-server (self-attestation): same merge gate, third artifact class](https://delimit.ai/reports/delimit-mcp-server-tdqs) — delimit-mcp-server (self-attestation)

See the full index at [delimit.ai/reports](https://delimit.ai/reports). For the schema and signing methodology behind every report, see [delimit.ai/methodology/mcp-attestation](https://delimit.ai/methodology/mcp-attestation).

---

Golden Path — your first 10 minutes

The API check, end to end: lint a spec, see exactly what breaks, classify the bump, settle the hard calls with the model providers you configure, and keep a locally signed receipt of what ran. Then keep the context that survives across sessions and models.

1. Install

npx delimit-cli scan          # discovery: finds your OpenAPI specs, frameworks, security issues, tests

npx delimit-cli init # wire up the merge-gate config (--preset strict | default | relaxed)

scan (delimit_scan) reports what Delimit can do for this repo. init (delimit_init) drops in the policy preset and merge-gate config. No account, no keys.

2. The merge gate, end to end

Lint the spec change (the gate). Baseline vs. proposed, with policy applied — one pass/fail verdict.

npx delimit-cli lint old.yaml new.yaml       # tool: delimit_lint

See exactly what breaks. Pure structural diff — added/removed/modified endpoints, schemas, params, no policy.

npx delimit-cli diff old.yaml new.yaml       # tool: delimit_diff

Classify the bump. Deterministic MAJOR/MINOR/PATCH/NONE — same input, same answer, every time.

delimit_semver       → MAJOR/MINOR/PATCH/NONE + next version string

delimit_impact → blast radius: scans your dependency manifest for downstream callers (informational)

Settle the hard calls. When the gate verdict is a judgment call, put it to multiple models and let them debate to consensus.

npx delimit-cli deliberate "Is dropping the deprecated v1 /users field a safe MINOR?"

tool: delimit_deliberate — uses your configured model CLIs or API keys

delimit setup gives the Codex MCP server a 30-minute tool timeout so a multi-round panel can return its completed transcript instead of being detached by the client while its models are still responding.

Capture the signed, replayable attestation. After a gate event (deploy / security / test / audit), record the evidence bundle and verify it any time.

delimit_evidence_collect   → signed evidence bundle for the audit trail        (Pro)

delimit_evidence_verify → confirm a bundle hasn't been tampered with (Pro) delimit_seal_verify → check a Delimit Seal receipt against its bundled (Free) Layer-0 constitution — offline-verifiable

Every receipt is offline-verifiable with npx delimit-cli seal-verify . Hosted receipt replay is not available yet.

3. Context that survives sessions and models

Decisions, constraints, and tasks persist across sessions. Continue in Claude Code, Codex, Cursor, Antigravity, GitHub Copilot, or Muse with the saved record.

Memory — persist and recall the *why*, not just the diff.

npx delimit-cli remember "v1 /users field is frozen until Q3 — downstream billing depends on it"

tool: delimit_memory_store

npx delimit-cli recall billing # local recall over your saved memories (Free)

For semantic recall by meaning across sessions, the assistant calls delimit_memory_search (Pro) directly.

Ledger — one task list, shared across every assistant and session.

delimit_ledger_add        → record a task/bug/feature/strategic item

delimit_ledger_context → session-start: top open items by priority (what's queued) delimit_ledger_done → close with a note (auto-captures a PR URL as ship proof)

That's the loop: gate the change, sign the proof, keep the context. Run it once on a real spec and you've used the whole merge gate.

Lead chat, Copilot CLI and Muse Code sessions

Install and authenticate the official client separately, under the same OS user who runs Delimit. Delimit does not install these clients, purchase a plan, log you in, or verify your subscription allowance. The interactive lead order is Claude → Codex → Antigravity → Muse → Copilot, subject to existing configuration, availability and native-client checks. This changes only lead chat, not review panels or background routing. Existing paid API fallback remains opt-in.

npx delimit-cli chat

npx delimit-cli chat --model muse npx delimit-cli chat --model copilot npx delimit-cli chat --project /path/to/repository --model muse

You can launch from your home directory without changing directories. That is an unscoped owner workspace: Delimit loads workspace instructions but does not infer a venture or restore the newest unrelated handoff. A stale active-venture pointer is not selected. Use --project to bind an existing Git repository; missing or invalid project paths fail before launching a model. Root/workspace exit does not claim a project-bound capture. Recover or save continuity against the actual task scope through existing Delimit tools.

Muse and Copilot are native interactive harnesses, not new model votes or automatic build-loop backends. An explicit --model muse or `--mode