Boundwrite

The Boundwrite API

Outbound messages, blog posts, SEO pages and LinkedIn posts, each bound to the facts you send. Write a message, review or edit long-form text one piece or 25 at a time, plan and post to LinkedIn, and turn trends into briefs. It is in beta and free to use.

Beta access

Request a key below. We review requests by hand and approve them within 4 hours, then email your key. Need it sooner? Email x@navyaai.com.

Limits during the beta

PriceFree during the beta.
Rate60 requests a minute per key.
EmailUp to 90 words in the body, subject of 1 to 7 words.
ReviewAny text up to 200,000 characters; request up to 1 MB.
BulkUp to 25 pieces per request, 4 MB; up to 80 distinct paragraphs edited.
Calendar1 to 12 weeks, 1 to 5 posts a week, at least 3 facts.
LinkedIn messageUp to 60 words and 500 characters.
LinkedIn connection noteUp to 280 characters.
FactsUp to 30 recipient facts and 30 sender facts, 600 characters each.
Request size64 KB.
TimeA request can take up to 100 seconds; most finish much sooner.

These may change while we are in beta. If you need more, email x@navyaai.com.

Authentication

Every request carries your key in a header: Authorization: Bearer YOUR_KEY. The base URL is https://api.boundwrite.com.

MCP clients don't need a key. They sign in with OAuth 2.1: authorization code with PKCE (S256), Client ID Metadata Documents or dynamic client registration, and tokens bound to https://mcp.boundwrite.com/mcp. Signing in is Sign in with LinkedIn, and during the beta only invited accounts get through. The discovery documents are at api.boundwrite.com/.well-known/oauth-authorization-server and mcp.boundwrite.com/.well-known/oauth-protected-resource. Endpoints: /authorize, /token, /register (every client is public, with no secret) and /revoke. The scope is boundwrite, and GET /v1/me shows who a token or key belongs to. A session lasts as long as your LinkedIn authorization, about 60 days, and then the client asks you to sign in again.

Write a message

POST /v1/outbound writes a message from the facts you send.

FieldTypeWhat it is
channelstringemail (default), linkedin_message or linkedin_connect.
recipient_factsstring[]What you know to be true about the recipient and their company. One fact per item.
sender_factsstring[]What is true about you and your offer. At least one.
signalstringOptional. Why you are writing now.
campaignstringOptional. Why you are writing to this group now, in your own words. Used once per sequence, as your view.
voice_samplesstring[]Optional. Up to 10 messages you wrote by hand. Used for your habits only, never for facts.
earlier_messagesstring[]Optional. Bodies already sent in this sequence or batch. The new message will not reuse their opener or restate them.
askstringThe one thing the message should ask for.
draftobjectOptional { subject, body }. Checked first, and returned unchanged if it already passes.

Request

curl https://api.boundwrite.com/v1/outbound \
  -H "Authorization: Bearer $BOUNDWRITE_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "channel": "email",
    "recipient_facts": [
      "Dana Reyes is head of support at Orbit Desk.",
      "Dana wrote a post saying first-reply time doubled after the last launch."
    ],
    "sender_facts": [
      "Kiln drafts first replies to support tickets from a team's own past answers.",
      "An agent reviews every Kiln draft before it is sent."
    ],
    "signal": "Orbit Desk posted 4 support roles this month.",
    "ask": "Ask whether a 15 minute call next week is worth it."
  }'

Response

{
  "message_id": "ee002d98af8da721e203614051ccc8ac",
  "channel": "email",
  "status": "written",
  "message": {
    "subject": "First reply time",
    "body": "Dana, you wrote that first-reply time doubled after the last launch, and Kiln drafts replies from your team's past answers for agents to review before sending. Would a 15 minute call next week be worth it?"
  },
  "claims": [
    { "claim": "You wrote that first-reply time doubled after the last launch.", "fact": "R2", "support": 0.97 },
    { "claim": "Kiln drafts replies from your team's past answers.", "fact": "S1", "support": 0.95 },
    { "claim": "Agents review drafts before sending.", "fact": "S2", "support": 0.83 }
  ],
  "issues": [],
  "questions": [
    "Paste 3 to 5 messages you wrote by hand (replies to a client or prospect work best) so the writer can match your habits.",
    "Why these people, why now? One sentence in your own words; it is used once, as your view."
  ],
  "judged": true,
  "attempts": 1
}

Fact ids are R1, R2… for recipient facts, S1, S2… for sender facts, G1 for the signal and C1 for your campaign line, in the order you sent them. Keep message_id if you want to match replies to messages later.

Result fieldWhat it is
statuswritten, unchanged or needs_review.
claimsEach factual claim, the fact id behind it, and support: the judge's probability that the fact backs it.
issuesWhy the message is needs_review: unsupported claims, generic lines, a vague opener, a repeated opener, rule breaks.
questionsWhat only you can supply: samples for voice, your reason for writing, a specific fact about the recipient. Answer and call again.
voice_match0 to 1, how closely the message matches your samples. Only with voice_samples.
judgedtrue when the judge ran. If it was unavailable, the claim list alone decided.

Your voice, and a sequence

Send voice_samples and the message is written in your habits: sentence length, how you open and close, how direct you are, whether you use an exclamation mark. Phrases and facts from the samples are never copied. Send campaign and your reason for writing is said once, as your view.

For a sequence, send each message already sent to that person in earlier_messages. The next one keeps a different opener, does not restate the earlier ones, and adds one thing. The same field works across a batch, so a hundred messages do not share one opening.

When the result has questions, something only you can supply is missing. Answer them and call again; the message is still returned so you can decide.

Check a draft

POST /v1/check takes the same body with a draft and judges it without rewriting. Use it on messages another tool wrote.

Review any text

POST /v1/review reviews a blog post, a page or a long email. It is deterministic, no model and no credits: a slop index from 0 to 100 (lower is better), six component scores (correctness, redundancy, genericness, evidence, structure, cadence) and findings with exact spans and fixes: filler phrases, paragraphs with no specifics, numbers with no source in the paragraph, superlatives with no basis, placeholders left in, repeated sentences and openers, dense sentences. It measures the writing, never who wrote it. For a short outbound message use /v1/check.

FieldTypeWhat it is
textstringThe draft, 40 to 200,000 characters. Markdown or plain text.
factsstring[]Optional. What the writer was given, one fact per item. Every number in the text that is not in the facts is reported as evidence.number_not_in_facts, which catches the figures a model makes up and an editor keeps.
originalstringOptional. The version this is a revision of. Adds fidelity: figures that changed or vanished and flipped negations are blockers; weakened hedges are review.
voice_samplesstring[]Optional. The author's own unedited writing, 2,000 words or more in 3 or more pieces. Adds voice: the draft's distance from the author, feature by feature.
voice_registerstringOptional. email (default), blog, docs or landing.
target_keywordstringOptional. For a page meant to rank: checks the H1 carries it.
page_typestringOptional. article, howto, comparison, roundup, glossary, local, pricing, service or casestudy. Sets strictness.
shapebooleanOptional. Also return the text after a paragraph-rhythm pass that changes no words.

Request

curl https://api.boundwrite.com/v1/review \
  -H "Authorization: Bearer $BOUNDWRITE_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "<the post, page or long email, markdown or plain text>",
    "facts": ["<optional: what the writer was given, one per item>"],
    "original": "<optional: the version before your edit>",
    "voice_samples": ["<optional: 2,000+ words of the author's own writing>"]
  }'

blockers are what must be fixed before publishing; review findings are judgement calls; info is advisory. Pass original after every edit, and after any paraphraser or humanizer, to see what moved. With voice_samples the response also carries voice_constraints, plain sentences about the author's habits with their numbers, for the prompt that writes the next draft. The engine is the one behind XeoRank's Writing Agent, from the same team.

Edit only the weakest sentences

POST /v1/improve takes text (and optional facts and voice_samples) and rewrites only the sentences that read most like generated text. Each rewrite must restructure the sentence, not swap synonyms, and keep every figure, negation, hedge, qualifier and name; short sentences are left alone. You get the new text and one patch per changed sentence with the reason.

Many pieces at once

POST /v1/bulk takes up to 25 documents. With mode: "check" no model is called: you get a review summary per piece, the safe filler fixes, the paragraphs that need a structural edit, and a batch report: pieces that open the same way, sentences shared across pieces, phrases most pieces use, and each piece's share of sentences found in the others.

With mode: "edit" (the default) the filler fixes are applied, and only flagged paragraphs go to a small model for two edits: deleting a sentence about the text rather than the subject, and merging two short sentences about one subject. An edit is kept only if no figure, negation, hedge, obligation, qualifier or name moved. A paragraph repeated across pieces is edited once. usage reports model calls, edits, reuses and deterministic fixes.

Request

curl https://api.boundwrite.com/v1/bulk \
  -H "Authorization: Bearer $BOUNDWRITE_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "mode": "check",
    "documents": [
      { "id": "post-1", "text": "<markdown or plain text>" },
      { "id": "post-2", "text": "<markdown or plain text>" }
    ]
  }'

Post to LinkedIn

Over MCP, signing in with LinkedIn is the connection: the same consent lets Boundwrite post to your own feed, and the LinkedIn token is kept sealed on our side against your account. There's nothing to copy.

With an API key, GET /v1/linkedin/connect returns a link instead. After LinkedIn's consent screen it shows a sealed connection string, which you send in the X-LinkedIn-Connection header. This path is deprecated and will go once OAuth covers every client.

POST /v1/linkedin/post takes text and optional facts, reviews the post, and without confirm: true returns a preview and posts nothing. Show the user the exact text, then call again with confirm: true. In beta.

Your profile and network, against your ICP

For personal profiles. PUT /api/v1/reach/icp takes a description of who you want to reach, in your own words, and returns it as functions, seniority, title words, target companies, exclusions and problems. Send icp as well to save an edited version as it is.

POST /api/v1/reach/profile takes your headline, about and optionally experience. It scores the profile out of 100 against the ICP and suggests rewrites. A suggestion is kept only if it uses facts already in your profile and passes the same checks.

POST /api/v1/reach/network takes the Connections.csv from LinkedIn's data export. It reports the share of your network in your ICP and the share of the last 90 days' new connections, which sets your level. It also shows who you've been adding instead and the ICP segments with the fewest people. Only titles, companies and dates are read. The file isn't kept; the report is, so the next upload shows the change. The dashboard is at /reach.

Your roadmap

GET /api/v1/roadmap/status returns your posting stage, this week's target and progress, your streak in weeks on cadence, what unlocks the next stage, this week's brief (theme, pillar, the format to use next, your audience and the next calendar slot) and the next actions in priority order. Your AI reads it through the MCP tool get_next_actions and drafts from the brief.

GET /api/v1/roadmap/context (MCP: plan_roadmap) gives your AI what it needs to plan twelve weeks: your ICP, profile gaps, network drift and the stage's cadence. It writes three pillars and a theme per week, and saves them with PUT /api/v1/roadmap (save_roadmap), which also takes timezone, max_per_week, goals and nudges ({morning, noon, evening}, each "HH:MM" or null for off; MCP: set_nudges). Posts Boundwrite publishes count on their own; log the ones you post yourself with POST /api/v1/roadmap/posts (log_post). GET /api/v1/roadmap/times says when to post next. The starting point is midweek, before work and at lunch, in the time of the ICP's main region (set with PUT /api/v1/reach/icp and {regions}). Once a slot has three posts, the person's own results take over. The dashboard is at /app, and it can send nudges to your browser.

Your own words, and cards

GET /api/v1/question gives the person one question a day, built from their ICP, pillars, this week's theme and stories from their niche. It costs no model tokens. Answers go to their story bank with POST /api/v1/stories (save_story), as they said them, and the roadmap brief hands the latest ones to their AI as own_words, so posts start from the person rather than the model.

POST /api/v1/cards (make_card) draws a post image from a design template, not an image model: number (big, line, foot), quote (text), list (title, items), notebook (title, lines) or carousel (title, slides). Each card is 1080 by 1350 in the person's own look: a design signature (one of the presets, a type pairing, a texture and their colours) given on their first card and kept apart from every other member's, so no two members post look-alike cards; the layout turns with each card. The style is set with PUT /api/v1/cards/style: name, tagline, handle, a font (slab, sans, mono or hand), three colours, and the signature's preset, type and texture. A card is refused if any figure on it isn't already in the person's own words. You get back signed image links to download and attach to a post.

A LinkedIn content calendar

POST /v1/calendar takes goals, an audience and at least three facts, and returns a roadmap with one theme per week tied to a goal, a dated slot per post in your timezone and a reviewed draft per slot. Drafts may claim only what is in the facts. A read-only share link comes back too.

POST /v1/calendar/{id}/slots edits a slot's text, at or mode, or approves it. A manual slot is posted by you; an auto slot is posted at its time, only after you approve that exact text. Editing the text withdraws the approval, the review runs again just before posting, and a slot more than six hours late is marked missed rather than posted late. GET /v1/calendar/{id} reports posted against planned per week, posts on time and the next slot. LinkedIn's own reach and engagement numbers arrive once LinkedIn approves our analytics access.

Request

curl https://api.boundwrite.com/v1/calendar \
  -H "Authorization: Bearer $BOUNDWRITE_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "goals": ["Book 5 audit calls from teams that just raised"],
    "audience": "CTOs of startups that recently raised and run their own inference",
    "facts": [
      "One audit took a team's monthly inference bill from $47K to $28K.",
      "Our Token Tax benchmark cut Llama 3 70B serving cost by 42%.",
      "The audit is free and written, for teams spending $20K+ a month on inference."
    ],
    "weeks": 3, "posts_per_week": 2, "timezone": "Asia/Kolkata", "time": "09:00"
  }'

Trends and opportunities

Onboard a brand with your key at POST https://boundwrite.com/api/v1/brands and add domains one at a time at /api/v1/brands/{id}/domains. Boundwrite reads up to 25 pages per domain, respecting robots.txt for BoundwriteBot, and learns your topics; you can add topics at /topics. It watches Hacker News, Google Trends and the news for them, and a match becomes an opportunity at /opportunities with the matched topics and an angle.

POST /api/v1/opportunities/{id} with action: "proceed" builds a brief from the story's own source, writes a draft from it that may not claim your experience (it marks [ADD: …] where only you can speak) and returns Boundwrite's review of the draft. Use accept or dismiss to sort the rest. Closed beta, five teams.

Request

curl https://boundwrite.com/api/v1/brands \
  -H "Authorization: Bearer $BOUNDWRITE_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "name": "NavyaAI", "domains": ["navyaai.com"] }'

Status

StatusMeaning
writtenA new message that passed every check. Safe to hand to your sequencer.
unchangedYour draft passed as it was.
needs_reviewNo version passed. You get the best attempt and the reasons in issues. Do not send it as is.

A message is written only if every claim cites a fact you supplied, it has no stock phrasing, it fits the channel, it asks exactly one question, and every number in it came from your facts. Two models check it: the writing model lists the claims it made, and a judge scores each claim against its fact, each sentence for generic filler, the opener for specificity and the draft against your samples. This reduces invented claims and generic copy but cannot rule them out.

Errors

Errors come back as {"error": "…"} with one of these codes.

CodeMeaning
401The key is missing or wrong.
422The request is invalid. The error says which field.
429Rate limit reached. Wait a minute.
502The writing model did not respond. Nothing was written; try again.

MCP server

For agents, the same operations are available as MCP tools at https://mcp.boundwrite.com/mcp (Streamable HTTP): write_outbound, check_outbound, review_content, improve_content, bulk_content, connect_linkedin, post_linkedin, create_content_calendar, get_content_calendar and update_calendar_slots.

Add the URL and your client opens a Boundwrite sign-in page. It names the client and what it may do, then sends you to LinkedIn. In Claude.ai, add it under Settings → Connectors as a custom connector. ChatGPT, Cursor and VS Code use the same URL.

Claude Code

claude mcp add --transport http boundwrite https://mcp.boundwrite.com/mcp

Scripts and older clients, with a key

claude mcp add --transport http boundwrite https://mcp.boundwrite.com/mcp \
  --header "Authorization: Bearer $BOUNDWRITE_KEY"

Help

Something broken or urgent? Email x@navyaai.com.