Liftli for developers and AI agents

Connect Liftli (one MCP URL)

Liftli is a single remote MCP server. The URL is the same across every client:

After connecting, the user signs in once with a secure WorkOS login. Liftli then extracts their writing voice and starts drafting in it. No tokens, no credits, no AI markup: it runs inside the user's existing AI subscription.

What an agent can do with Liftli

Liftli is not a fixed set of buttons. It is an agent surface, so the ceiling is what you can ask an agent to do. A few real workflows:

• From a voice memo to a scheduled week. "Take my last voice note and draft and schedule five posts across LinkedIn and X." Liftli plans against the saved strategy, drafts in the user's voice, critiques each draft, and queues them behind one-tap approval.
• From real work to post ideas. Point it at call transcripts, GitHub activity, or a chat, and it mines them for post "seeds" grounded in what actually happened, instead of a blank prompt box.
• Compose with the user's other tools. Because it lives inside the assistant, you can pull ideas straight from the user's connected transcription or notes tool and hand them to Liftli in one step, no export, no dashboard.
• Close the loop. After a post ships, Liftli scores who engaged into a ranked warm-lead list and recalibrates next week's plan on what worked.

Full capability list and the tool catalog are in llms.txt and the expanded llms-full.txt.

Free skills your agent can run locally

Separate from the MCP, all 28 of Liftli's free tools ship as installable skills. The agent's own model runs the distilled methodology locally, no account required:

npx skills add liftli-ai/skills

npx skills add liftli-ai/skills --skill linkedin-hook-generator

Browse all 28 at liftli.ai/skills, or read the machine-readable manifests: /.well-known/skills.json and /.well-known/agent-skills/index.json. Source: github.com/liftli-ai/skills.

Why agent-native, and why it is ban-safe

Most content tools are dashboards you open, or browser extensions that act on your logged-in session. LinkedIn's User Agreement prohibits scraping and extension-driven automation, and it enforces that. Liftli sits on the other side of that line: it publishes and schedules through the platforms' official APIs, connected only when the user explicitly links an account, and it never runs a bot action on a profile. Nothing publishes without the user's one-tap approval.

The agent-native shape is also the reason the tool inherits the user's context for free. It is already inside the assistant that holds the user's calls, code, and thinking. That is the layer where content ideas actually happen, and where an agent can carry them all the way to published.

Authentication

Two surfaces, two answers.

The free micro-tools API needs no credentials at all. GET /health, GET /version and everything under /api/v1/tools/ are open to anonymous callers. There is no key to request, no form to fill in, and no sales step — an agent can call them on first contact. They are rate limited per caller per day, and every successful generation tells you how many you have left.

The MCP endpoint uses OAuth 2.1 with PKCE. Point your client at https://mcp.liftli.ai/mcp. An unauthenticated initialize returns 401 with a WWW-Authenticate header naming the protected-resource metadata, exactly as RFC 9728 describes:

WWW-Authenticate: Bearer resource_metadata="https://mcp.liftli.ai/.well-known/oauth-protected-resource"

That document names https://auth.liftli.ai as the authorization server. It supports dynamic client registration, so an agent can register itself and complete the flow without a human provisioning credentials first. Authorization-code and device-code grants are both supported; PKCE is S256.

Request least privilege. The available scopes are openid (identify the account), profile, email, and offline_access (refresh without the user present). openid alone is enough to identify a signed-in Liftli account.

REST endpoints and example requests

The full machine-readable contract is /openapi.json (OpenAPI 3.1, with typed schemas, unique operationIds and declared OAuth scopes). Base URL: https://app.liftli.ai.

Discover which tools are available — call this first; it returns the valid values for the tool field.

curl https://app.liftli.ai/api/v1/tools/health

{"enabled": true,
 "tools": ["about-generator", "carousel-outline", "comment-generator", "content-ideas",
           "contrarian-angles", "devlog-to-post", "headline-generator", "hook-generator",
           "hot-take-check", "linkedin-to-x", "meeting-to-post", "poll-generator",
           "post-generator", "post-rewriter"]}

Run one generation. inputs is a flat map of strings; the accepted keys depend on the tool.

curl -X POST https://app.liftli.ai/api/v1/tools/generate \
  -H 'Content-Type: application/json' \
  -d '{"tool": "hook-generator",
       "inputs": {"topic": "why most AI posts sound the same",
                  "audience": "B2B founders"}}'

{"result": "Everyone wants the AI. Nobody wants the cleanup.",
 "remaining_today": 4}

Check service status and client compatibility.

curl https://app.liftli.ai/version

{"version": "1.4.6", "min_client_version": "1.0.0",
 "deprecated_endpoints": ["/log-reply"], "status": "stable"}

List the MCP tools (needs a bearer token from the OAuth flow above).

curl -X POST https://mcp.liftli.ai/mcp \
  -H 'Authorization: Bearer <token>' \
  -H 'Content-Type: application/json' \
  -d '{"jsonrpc": "2.0", "id": 1, "method": "tools/list", "params": {}}'

Errors, rate limits and quotas

Every error is JSON — there are no HTML error pages on the API. A validation failure returns 422 with a detail array naming the offending field, so an agent can correct the call without guessing:

curl -X POST https://app.liftli.ai/api/v1/tools/generate \
  -H 'Content-Type: application/json' -d '{}'

HTTP/1.1 422 Unprocessable Entity
Content-Type: application/json

{"detail": [{"type": "missing", "loc": ["body", "tool"], "msg": "Field required"}]}

Anything else returns the same shape with a string detail — for example a missing route gives 404 and {"detail": "Not Found"}.

The micro-tools API is quota'd per caller per day rather than per second. Read remaining_today on each successful response and stop when it reaches zero; a call past the quota returns 429. Treat 429 as retryable only after the quota window resets — retrying immediately will not succeed.

Versioning and deprecation policy

The API is versioned in the URL path: the current surface is /api/v1/. A new major version gets a new path prefix; v1 keeps working while it does.

Within a version we add fields but do not remove or repurpose them, and we do not change the type of an existing field. Treat unknown fields as forwards-compatible and ignore them rather than failing.

How you learn something is going away. GET /version is the machine-readable source of truth. It returns min_client_version (the oldest client still accepted) and deprecated_endpoints — an array of paths that still respond today but are scheduled for removal. Poll it, or check it at startup, and migrate off anything listed there.

curl https://app.liftli.ai/version

{"version": "1.4.6", "min_client_version": "1.0.0",
 "deprecated_endpoints": ["/log-reply"], "status": "stable"}

A deprecated endpoint stays available for at least 90 days after it first appears in deprecated_endpoints. Breaking changes are announced on the blog before they ship. If you are building against Liftli and need a longer window, say so — it is one person, and that conversation is easy to have.

Machine-readable resources

• /llms.txt — concise index for LLMs: what Liftli is, install, capabilities, tools, pricing.
• /llms-full.txt — the expanded single-file reference (capabilities, full skills catalog, glossary, FAQs).
• /ai-agents.md — a short directive for an AI agent helping a user grow on LinkedIn, X or Substack.
• /.well-known/skills.json and /.well-known/agent-skills/index.json — the skills manifests.
• /sitemap.xml — every indexable page.