</> 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.
Using the Delimit Claude plugin?
The plugin needs Claude Code or Cowork (it does not run on claude.ai on the web).
The first thing to try, as a plain prompt in Claude Code:
And the next two:
"record this decision: ..."
"hand off this session"
Want it automatic? Run npm install -g delimit-cli then delimit setup: when a Claude Code session starts, Claude gets your last working context, open tasks and last handoff without you asking.
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
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 "<question>" 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.
$ 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
delimit check --base main
delimit check --fix
delimit check --record
--record writes a content-pinned record of the check to .delimit/records/check-<ts>.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:
See the full index at delimit.ai/reports. For the schema and signing methodology behind every report, see 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
npx delimit-cli init
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
See exactly what breaks. Pure structural diff — added/removed/modified endpoints, schemas, params, no policy.
npx delimit-cli diff old.yaml new.yaml
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?"
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 <receipt.json>. 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"
npx delimit-cli recall billing
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 --model copilot
stays on that selection without probing or silently switching providers.
In the default lead chain a failed native launch can advance to the next lead;
an intentional interrupt exits. No successful handoff is inferred from a failed
or unavailable capture. Native permission approvals remain active. Copilot launches
with automatic updates and remote/export features disabled for this session.
Muse preparation: first verify that your account permits Standard
muse-spark-1.3 on the intended billing route. Back up your native
$XDG_CONFIG_HOME/muse/settings.json (or ~/.config/muse/settings.json when
XDG_CONFIG_HOME is unset), and merge these fields into its existing JSON object;
preserve unrelated settings and authentication:
{
"model": "muse-spark-1.3",
"context": {
"foreign_personal_rules": false,
"foreign_personal_skills": false
}
}
The explicit Muse launch requires that selection and exclusion of foreign
personal rules/skills; it does not change native settings for you. A configured
provider must be meta, and a configured endpoint must be
https://api.meta.ai/v1. META_API_KEY or MODEL_API_KEY environment overrides
block this launch until you resolve the intended billing route. Do not switch
credentials or models merely to bypass that check. Standard selection is not
proof of zero retention, subscription billing, or private-data clearance.
Current binary discovery supports Copilot in ~/.local/bin, /usr/local/bin,
or /usr/bin, and Muse's versioned ~/.local/bin/muse-bin-<version>-R<revision>
distribution. Muse uses that executable directly, not its updater launcher.
Other installation layouts are not automatically discovered. The native Muse
context settings were checked with Muse Code 1.2.1; recheck compatibility when
upgrading the client.
Delimit supplies a compact project-bound startup instruction, not an automatic
import of your ledger, memory, or complete conversation. Optional local guidance
can live in $DELIMIT_HOME/harness-bootstrap.md (default
~/.delimit/harness-bootstrap.md); the combined startup text must fit within
16 KiB. Project AGENTS.md/CLAUDE.md files at or above 64 KiB are rejected
instead of silently truncated. Recover and save a scoped handoff through your
available Delimit tools. These launchers do not automatically register a
native MCP connection or prove cross-harness continuity.
When both project instruction files exist, the startup packet lists both files
with their sizes and hashes and tells the harness to read them completely,
including guidance hidden by native file precedence. It does not copy their
contents into the startup packet, silently merge files, or override the existing
instruction hierarchy. User-global guidance remains opt-in through the local
bootstrap file; Delimit never copies a different client's private rules by default.
This manifest is a loading instruction, not proof that the model performed every
read. If required guidance is inaccessible or conflicts, resolve that gap before
allowing edits. Oversized native project files still fail closed without rewriting
the originals.
Optional command shims can be installed with delimit setup --harness-shims.
Ordinary setup only refreshes already-managed Muse/Copilot shims; it does not
newly shadow your native commands. Custom shims are preserved. With the opt-in,
a conflicting custom file stops installation of both additional shims.
For existing launcher-managed continuity, clean exit finalizes only a bound
handoff. If that is unavailable, Delimit reports the specific reason and writes
a private local diagnostic receipt under its sessions/ directory. That receipt
is not a conversation backup. Existing handoffs are not replaced by empty ones;
resume the exact native session and explicitly save a project-bound handoff.
Think and Build
delimit think runs a multi-model deliberation using available configured providers. delimit build starts a detached planning worker that inspects the repository and writes briefs; it does not execute ledger tasks. delimit vault stores secrets as plaintext JSON in owner-only (0600) files.
Works across any configuration, from a single model on a budget to a full panel.
Try it in 2 minutes
npx delimit-cli doctor
npx delimit-cli status
npx delimit-cli simulate
npx delimit-cli scan
npx delimit-cli try owner/repo
No API keys. No account. No config files.
Pick your first win
Protect my API — catch breaking changes before merge:
Watch for drift — detect spec changes without review:
Run PR copilot — governance gates on every pull request:
- uses: delimit-ai/delimit-action@v1
with:
spec: api/openapi.yaml
What's New
Gate every AI-assisted invocation. Ship the receipts.
delimit wrap — pipe claude -p, cursor, aider, codex, or any AI-assisted CLI through a signed governance gate. Snapshots the git diff before/after, runs lint + tests, HMAC-signs an att_* attestation, and writes the receipt locally for offline verification. Advisory by default; --enforce blocks CI on policy violations; --max-time <s> is a kill switch that tags the attestation as a liability_incident and prints a cross-model handoff command.
delimit trust-page — renders a directory of attestations into a static HTML trust page + JSON Feed 1.1 feed. Single file, no framework, offline-renderable. Deploy anywhere.
delimit ai-sbom — aggregates attestations into a CycloneDX 1.6 bill-of-materials with AI-specific fields (detected models per vendor, tool-call surface, policy gate counts). Pipe straight into procurement.
- Cross-model by construction —
wrap uses the same local attestation schema across supported command-line producers. Delimit keeps session context across Claude Code, Codex, Cursor, Antigravity, GitHub Copilot, and Muse.
delimit wrap -- claude -p "add tests for payments"
delimit wrap --max-time 60 -- cursor edit "refactor auth middleware"
delimit trust-page -o ./trust
delimit ai-sbom -o ./ai-sbom.json
Earlier releases
The highest state of AI governance — earlier features still active.
delimit doctor -- checks your local setup and suggests fixes. Supports --ci for pipelines and --fix for supported repairs.
delimit simulate -- policy dry-run. See what would be blocked before you commit. The terraform plan for API governance.
delimit status -- visual terminal dashboard. Policy, specs, hooks, CI, MCP, models, memory, ledger, evidence, git branch. --watch for live refresh.
delimit report -- governance report. --since 7d --format md|html|json. Audit-friendly output for PRs and compliance.
- Memory hardening -- SHA-256 integrity hash + source model tag on every
remember. Cross-model trust, verified on every recall.
- Tag-based publishing -- automated gateway sync, no more version drift between source and npm bundle.
Multi-Model Deliberation
Run your question through 4 AI models simultaneously. They debate each other until unanimous agreement.
delimit deliberate "Should we build rate limiting in-house or use a managed service?"
Round 1 (independent):
Claude: Build in-house. Redis sliding window is 50 lines.
Gemini: Build. You already have Redis.
Codex: Agree — but add circuit breaker for Redis failures.
Grok: Build. Managed service costs $200/mo for 50 lines of code.
Round 2 (deliberation):
All models: AGREE
UNANIMOUS CONSENSUS (2 rounds, confidence 94/100)
Build rate limiting in-house with Redis + circuit breaker.
Deliberations use your configured model CLIs or API keys. Provider availability depends on your setup.
v4.18
- MCP Registry record published on every release -- the official registry entry now tracks the npm version automatically. Publisher pinned and checksum-verified; the tag gate asserts that both
server.json version fields match the package version.
- Tool descriptions -- 16 previously undocumented parameters documented from their signatures; 26 Pro-gated tools state the prerequisite and the unlicensed-call response.
- Release guards -- bundle-classification guard no longer flakes on SIGPIPE.
v4.17
- Cross-model continuity on fresh installs --
delimit_soul_capture / delimit_revive work without any internal modules via the public session_continuity module; captured souls carry deterministic provenance (venture, transcript identity, capture key) for reliable resume across model switches. Existing soul files load unchanged.
- Security audit -- fewer false positives on test fixtures and documentation dummies; suppressed findings are reported, never silently dropped.
v4.16
- Fail-closed bundle allowlist -- the npm package ships only explicitly reviewed gateway files; two CI guards block any unclassified or non-allowlisted file from entering the tarball.
- Session-end auto-capture installs in the settings shape Claude Code expects, so end-of-session handoff fires instead of being silently dropped.
- Always-on commit-author audit workflow catches identity drift on the repo.
v4.1
- TUI -- terminal-native Ventures panel, real
delimit think and delimit build commands
- Security hardening -- notify.py stubbed in npm, axios pinned against supply chain attacks
- Deliberation providers -- deliberations run on the model CLIs or API keys you configure (
delimit models); no hosted runs are included
- Zero-config onboarding -- auto-detect framework, scan, and first evidence in one command
- Assistant permissions --
delimit setup configures MCP access for Claude Code, Codex, Cursor, and Antigravity and sets approval defaults for some clients; review these settings for your environment
v4.0
- Toolcard Delta Cache -- SHA256 schema hashing, delta-only transmission, saves tokens
- Session Phoenix -- cross-model session resurrection with soul capture
- Handoff Receipts -- structured acknowledgment protocol between agents
- Cross-Model Audit -- 3 lenses (security, correctness, governance) with deterministic synthesis
- 4-model deliberation -- Claude + Grok + Gemini + Codex debate until consensus
- Universal Swarm Triggers -- "Think and Build", "Keep building", "Ask Delimit"
- Full governance toolkit -- lint, diff, policy, evidence, drift, attestation, and swarm orchestration exposed as MCP tools and CLI subcommands
GitHub Action
Zero-config -- auto-detects your OpenAPI spec:
- uses: delimit-ai/delimit-action@v1
Or with full configuration:
name: API Contract Check
on: pull_request
jobs:
delimit:
runs-on: ubuntu-latest
permissions:
pull-requests: write
steps:
- uses: actions/checkout@v4
with:
fetch-depth: 0
- uses: delimit-ai/delimit-action@v1
with:
spec: api/openapi.yaml
That's it. Delimit auto-fetches the base branch spec, diffs it, and posts a PR comment with breaking changes, semver classification, migration guides, and governance gate results.
View on GitHub Marketplace | See a live demo (23 breaking changes)
Breaking Changes Detected
| Change | Path | Severity |
|---|
| endpoint_removed | DELETE /pets/{petId} | error |
| type_changed | /pets:GET:200[].id (string -> integer) | warning |
| enum_value_removed | /pets:GET:200[].status | warning |
Semver: MAJOR (1.0.0 -> 2.0.0)
Migration Guide: 3 steps to update your integration
Governance Gates
| Gate | Status | Chain |
|---|
| API Lint | Pass/Fail | lint -> semver -> gov_evaluate |
| Policy Compliance | Pass/Fail | policy -> evidence_collect |
| Security Audit | Pass | security_audit -> evidence_collect |
| Deploy Readiness | Ready/Blocked | deploy_plan -> security_audit |
Adopt with minimum privilege
You don't have to trust a large tool surface on day one. The safe on-ramp:
Phase 1 — read-only governance (free, no account). Start with the tools that
only read your repo and write reports: delimit_lint, delimit_diff,
delimit_semver, delimit_policy, delimit_explain, delimit_scan, and
delimit_seal_verify. If your MCP client supports per-tool allowlists, grant
exactly those. Nothing in this set executes, deploys, or posts anywhere.
Phase 2 — opt into side effects deliberately. Tools that write evidence
bundles, open PR comments, or run deploys (delimit_security_audit,
delimit_deploy_*, agent orchestration) are tier-gated; enable them once
phase 1 has earned its keep in your CI.
Pin the Action to a commit SHA. @v1 is a floating tag. For
supply-chain-sensitive pipelines, pin the exact commit and bump on review:
- uses: delimit-ai/delimit-action@<commit-sha>
Keep BYOK keys out of repo and dotfile config. If you bring your own model
keys for deliberation, store them with delimit_secret_store (kept locally in
owner-only 0600 files, base64-encoded, not encrypted; access-logged via
delimit_secret_access_log) rather than in dotfiles.
Our own releases ship under the same discipline: every release carries a
signed, replayable Seal receipt (see the latest
release assets —
verify with npx delimit-cli seal-verify <receipt.json>), plus SLSA
provenance on npm.
CLI commands
npx delimit-cli scan
npx delimit-cli pr owner/repo#123
npx delimit-cli quickstart
npx delimit-cli try
npx delimit-cli demo
npx delimit-cli init
npx delimit-cli init --preset strict
npx delimit-cli setup
npx delimit-cli setup --dry-run
npx delimit-cli lint api/openapi.yaml
npx delimit-cli diff old.yaml new.yaml
npx delimit-cli explain old.yaml new.yaml
npx delimit-cli check
npx delimit-cli check --staged --fix
npx delimit-cli hooks install
npx delimit-cli hooks install --pre-push
npx delimit-cli ci
npx delimit-cli ci --strict --dry-run
npx delimit-cli remember "Redis uses JWT 15min"
npx delimit-cli recall redis
npx delimit-cli recall
npx delimit-cli recall --tag deploy --all
npx delimit-cli recall --export
npx delimit-cli forget abc123
npx delimit-cli models
npx delimit-cli models --status
npx delimit-cli status
npx delimit-cli doctor
npx delimit-cli uninstall --dry-run
npx delimit-cli wrap -- claude -p "..."
npx delimit-cli wrap --max-time 60 -- codex "..."# With kill switch + handoff on timeout
npx delimit-cli trust-page -o ./trust
npx delimit-cli ai-sbom -o ./ai-sbom.json
When installed into your AI coding assistant, Delimit provides tools across two tiers:
Free (no account needed)
- API governance -- lint, diff, policy enforcement, semver classification
- Persistent ledger -- track tasks across sessions, shared between all AI assistants
- Zero-spec extraction -- generate OpenAPI specs from FastAPI, Express, or NestJS source
- Project scan -- auto-detect specs, frameworks, security issues, and tests
- Quickstart -- guided first-run that proves value in 60 seconds
Pro
- Multi-model deliberation -- the models you configure debate until they agree; provider availability depends on your setup
- Security audit -- dependency scanning, secret detection, SAST analysis
- Test verification -- confirms tests ran, measures coverage, generates new tests
- Memory & vault -- persistent context and local plaintext secret files across sessions
- Evidence collection -- governance audit trail for compliance
- Deploy pipeline -- governed build, publish, and rollback
- OS layer -- agent identity, execution plans, approval gates
What It Detects
Delimit checks OpenAPI and JSON Schema changes using deterministic rules for the cases each surface supports. Examples include removed endpoints, newly required parameters, and changed response fields. Review the result alongside your contract; rule coverage differs between the CLI and GitHub Action.
Policy presets
npx delimit-cli init --preset strict
npx delimit-cli init --preset default
npx delimit-cli init --preset relaxed
Or write custom rules in .delimit/policies.yml:
rules:
- id: freeze_v1
name: Freeze V1 API
change_types: [endpoint_removed, method_removed, field_removed]
severity: error
action: forbid
conditions:
path_pattern: "^/v1/.*"
message: "V1 API is frozen. Changes must be made in V2."
- OpenAPI 3.0 and 3.1
- Swagger 2.0
- YAML and JSON
FAQ
How does this compare to Obsidian Mind?
Obsidian Mind is a great Obsidian vault template for Claude Code users who want persistent memory via markdown files. Delimit takes a different approach: it keeps a shared record across Claude Code, Codex, Cursor, Antigravity, GitHub Copilot, and Muse. Your memory, ledger, and governance travel with you when you switch models. Delimit also adds API governance (breaking-change detection for OpenAPI specs), CI gates, git hooks, and policy enforcement that Obsidian Mind doesn't cover. Use Obsidian Mind if you're all-in on Claude + Obsidian. Use Delimit if you switch between models or need governance.
Does this work without Claude Code?
Yes. Delimit supports Claude Code, Codex, Cursor, Antigravity, GitHub Copilot, and Muse. The remember/recall commands work standalone. The MCP server integrates with clients that support the Model Context Protocol.
Is this free?
Delimit lists Free, Pro, and Enterprise plans. Multi-model deliberation uses your configured model CLIs or API keys. See current plan details.
Local data and network calls
Delimit stores its ledger, records, and local attestations on your machine. Some setup and session paths make network requests by default:
- Setup checks for a newer CLI release and can download the Pro module tarball.
- The Claude Code SessionStart hook and installed shims can check for and install updates.
- License activation contacts the license provider and sends your customer email, license ID, and machine hash to delimit.ai.
- Multi-model deliberation sends your prompt and relevant context to the model providers you configure, through your CLI logins or API keys.
gateway/ai/supabase_sync.py is not bundled with this package. Webhook notifications are sent only when DELIMIT_WEBHOOK_URL is configured.
Links
MIT License