Changelog
Dated NeuroAPI releases, newest first, with breaking changes, new limits, and billing behavior in the public API.
Current build
- API version:
v1 - Build:
0.1.19
The same values come back in the X-API-Version and X-Build headers on
every API response.
v0.1.19
2026-11-16
- No API changes - this build only changes how NeuroAPI is operated. Requests, responses, limits and billing behave exactly as in v0.1.18.
v0.1.18
2026-11-14
- Live token data in more answers - token metrics, technical
analysis, trending coins and pools, new and recently added coins,
top holders, smart-money flows, buy and sell pressure and token risk
checks were unavailable to
POST /agent/askand the MCPasktool. The agent now uses them in every mode. - Coded errors instead of
500- a non-streamingPOST /agent/askthat hits its step limit before it produces youroutput_schemanow returns422 agent_budget_exceeded, and unusable model output returns502 agent_model_error. Neither is billed. Errors says when to retry each. - Long runs finish with an answer - a
POST /agent/askrun close to its step limit now wraps up with the data it already gathered instead of ending in an error. - No more false
429 too_many_concurrent_runs- a run stopped part-way, for example during a deploy, could keep counting against your concurrent-run limit. A stopped run now frees its slot within 3 minutes.
v0.1.17
2026-11-12
- No more intermittent
500onPOST /agent/ask- some calls, mostlysmartandmaxruns that drew extra quota, could return500after the answer had already been produced. These calls now return the answer as normal. X-Quota-Remainingis always a whole number - the header now always reports the number of requests left in your quota.- You were not billed for the failed calls - every call that hit this
500was refunded to your quota, so no action is needed on your side.
v0.1.16
2026-11-10
- Retry the right
429- Rate limits shows how to tellrate_limited, which clears afterRetry-After, fromquota_exceeded, which clears atX-Quota-Reset. The retry samples no longer retry a spent quota. - Accurate error reference - a malformed request body returns
422, and Errors now says so. It also covers the plain-text413, the504agent_timeout, and which statuses are safe to retry. - Failed streams are documented - a stream that fails ends with an
errorevent and no[DONE], and the call isn't billed. - Sample timeouts match the server - the code samples now wait 300 seconds, the limit for a non-streaming run, so they no longer give up on answers that are still running.
- Concurrent runs on Pricing - Pricing now lists how many runs each plan can have at once: Starter 2, Pro 5, Enterprise 10.
- Clearer API reference - every error response says what to do next,
and
modelists its cost and the plans that include it. MCP clients see the same mode details. - Fixed code samples - two broken samples now run as written.
v0.1.15
2026-11-03
- Compare, MCP, and Agents - new public pages compare NeuroAPI with market data and AI search APIs, explain how to connect MCP clients, and describe what the agent can do. All three are linked from the site footer.
- MCP rate limits - MCP requests follow the same rate-limit and auth rules as the HTTP API. See Rate limits.
- Pricing on mobile - all plans can be compared on small screens.
- Security disclosures - report vulnerabilities to
[email protected]as described on the Security page. - Docs navigation - missing pages show a not-found state.
- API key creation dialog - Escape dismisses the copy hint before closing the dialog.
v0.1.14
2026-10-30
- Choose your quota emails - get them at 50%, 80%, 90%, or 100%, or turn them off, under Account → Notifications. You now also get an email when the quota runs out. See Quota emails.
- Edit API keys - rename a key or change its expiry without rotating the secret.
- Revoke all keys at once - from Account → Security, after confirming your password. See Revoke every key at once.
- New home page, About, and FAQ - neuroapi.neurobro.ai shows sample answers, pricing, the AI tools it works with, and common questions before you sign in.
- Use every unit of your quota - a
modethat costs more than you have left now runs on what's left instead of returning402quota_exceeded.402now only means a payment failed. - Dashboard updates - a new Account page, a softer dark theme, and
keyboard shortcuts (
⌘/orCtrl+/lists them). - Expired keys show as Expired in the dashboard. They used to show as Active.
- The sidebar shortcut no longer blocks the browser's Save - it moved
from
⌘Sto⌘B. - Terms and Privacy open without signing in.
- Security patches across the API and the docs.
v0.1.13
2026-10-28
-
Breaking: new streaming event shape - with
stream=true, every event is now{"type": ..., "data": {"answer": ..., "usage": ...}}.typeisstatus_updatewhile the agent works andcompleteat the end. Themetadataobject, tool-progress events, and internal status details are gone. The answer text is still atdata.answer.chunk.content. If you readmetadata.message_type, readtypeinstead.Usage arrives once near the end with the same fields as a non-streaming response, so it has
cost_unitsand no longer hastotal_tokens. See Streaming. -
A limit on runs in flight at once - each account can now have 2 (Starter), 5 (Pro), or 10 (Enterprise)
POST /agent/askruns going at the same time, MCP calls included. One more returns429too_many_concurrent_runswithRetry-Afterand isn't billed. See Rate limits. -
Long runs stop after 5 minutes - a non-streaming run that takes longer than 300 seconds now returns
504agent_timeoutand isn't billed. Stream long research withstream=true. -
Request bodies are capped at 8 MiB - larger bodies get
413with a plain-text body. That is well above the largest valid request. -
message_historyroles - onlyuserandassistantturns reach the agent.system,developer,tool, andfunctionturns are accepted and ignored, so put instructions insystem_promptinstead. Any other role now returns422instead of500. -
Expired API keys stop working - a key whose
expires_athas passed now gets401. Before, it kept working. -
Streams that fail after the full answer are billed - if the final answer was delivered and something fails afterward, the stream ends with
[DONE]and counts as a run. Failures before the answer is complete are still free. -
MCP answers on
/api/v1/mcpdirectly - the path without a trailing slash no longer redirects, which some MCP clients didn't follow. -
Agent runs continue after tool-call errors - a malformed tool call or a missing trading symbol no longer ends the run. Separately,
mode=maxuses a stronger model for long research.
v0.1.12
2026-10-23
- MCP rejects API keys in the URL - requests to
/api/v1/mcpthat carry an API key in the query string now return400api_key_in_query. Proxies and logging tools often record full URLs, so send the key only in theX-API-KeyorAuthorization: Bearerheader. If a key has ever been sent in a URL, rotate it from the dashboard.
v0.1.11
2026-10-16
- Structured output is checked against your schema - when you send
output_schema, the object you get back is checked for field types, required fields, allowed values, and number limits. If the agent gets something wrong, it corrects it before replying. If it still can't, the request fails and you aren't charged. 422errors for unsupported schemas - schemas are read as JSON Schema 2020-12. These now return422with the exact location of the problem: regular expressions (pattern,patternProperties), dynamic references ($dynamicRef,$recursiveRef), references that don't point to a definition in$defs, and schemas nested deeper than 32 levels or made of more than 500 parts.nullinstead of made-up values - if a field allowsnulland the agent can't find the value, you getnullinstead of a stand-in like0. Make such fields nullable. See Designing the schema.- Backup model on
fastandsmart- if the AI model cuts a response off halfway, the request now switches to a backup model instead of failing.
v0.1.10
2026-10-15
- Breaking: a failed payment returns
402, not403- when your subscription's latest payment fails,/api/v1requests return402payment_requiredinstead of403no_active_subscription. Access returns as soon as a payment succeeds. If you handle403no_active_subscription, handle402too. See Billing. Idempotency-Keyno longer replays402,403, or429- a retry with the same key now runs again once you've paid, upgraded, or waited out the quota or rate limit, instead of returning the stored rejection. Stored replies now expire after 24 hours, as documented. See Idempotency.mode=fastandmode=smartloop less on long conversations - fewer requests run long and return a repetitive or cut-off answer.- Subscriptions from restricted countries are refunded - if a
subscription's billing address or card turns out to be from a country
NeuroAPI can't serve, it is canceled and the payment refunded. Requests then
return
403no_active_subscription. - Checkout open again in Syria - following changes to U.S. sanctions policy. Other listed countries remain restricted at checkout and on billing details.
v0.1.9
2026-10-01
- Checkout compliance by location - before paid checkout starts, NeuroAPI checks whether your connection appears to be from a restricted country. Local development is unaffected, and checkout still works when your country can't be determined.
- Billing details verified for compliance - your billing address and card country are also checked when a subscription is created or updated, so a restricted region can't complete checkout by skipping the initial location check alone.
v0.1.8
2026-09-30
- Bug fixes - minor fixes and stability improvements.
v0.1.7
2026-08-21
X-Request-Idon every response - the header was set only on a successfulPOST /agent/ask, so a401,403,422,429, or aGET /healthcall came back without one, even though those are the responses you need to report. Every response now carries an id, and a replayedIdempotency-Keyreturns the original request's id instead of a new one.- Every error response now has a documented schema - error bodies come in
three shapes: coded (
{"detail": {"code", "message"}}), plain ({"detail": "..."}), and the list-shaped422. All three are now in the API reference, and the Errors guide lists everycodewith the status it arrives on and what clears it. - New Idempotency guide - covers the 24-hour
replay window, the
Idempotent-Replayedheader, why a replay is never billed twice, and the cases that don't replay. - Quota and rate-limit headers documented -
X-RateLimit-*are now declared on the success response as well as under429, so you can read them before you're throttled. TheX-Quota-*headers reporting your remaining plan budget are now covered in Billing. mode=fastandmode=smartmigrated to a current-generation model - the model previously serving both modes has been retired by its provider. Both now run on its successor and fall back to an independent second provider, so an incident at the first provider no longer surfaces as a5xx.mode=maxruns on a newer primary model with a three-deep chain behind it.mode=maxtrims long multi-turn history - a largemessage_historycould previously outgrow the model's context window and surface an upstream400mid-run. The history you submit is now authoritative and is trimmed locally ahead of each step, so a long conversation no longer hits that400.- On-chain research -
POST /agent/askcan now screen a token for scam flags, mint/freeze authority, and top-holder concentration, read buy/sell pressure across 5m-24h windows, rank a token's most profitable traders, and surface newly launched tokens filtered by age and liquidity.
v0.1.6
2026-07-14
- Call the agent over MCP - NeuroAPI now supports the Model Context
Protocol. Point any MCP client at
https://api.neurobro.ai/api/v1/mcp(remote, Streamable HTTP) with yourX-API-Key, and theask_neuroapitool runs the same single-turn research asPOST /agent/ask, with the same rate limits, metering, and billing. See the MCP guide to connect.
v0.1.5
2026-06-28
- Trade-setup confidence is tied to risk/reward - when
POST /agent/askincludes a trade setup, its 1-10 confidence is anchored to the setup's own risk/reward and structure instead of being a free-form guess. The same setup now scores consistently, and a high number can't sit on a weak risk/reward, so you can set a threshold on it. The score rates the setup and can't tell you how a trade will end. - Tool use on
mode=max- changes to howmode=maxselects and calls tools mean requests that need live data (market metrics, technical analysis, trade setups) stall less often on tool selection. - Searchable docs - the documentation site now has full-text search. Press ⌘K (or Ctrl-K) or use the search bar to jump to any endpoint, guide, or field.
v0.1.4
2026-06-03
-
Structured output with a JSON Schema -
POST /agent/asknow accepts an optionaloutput_schema(a JSON Schema with an object root) and returns the validated object on a newoutputfield.answeris thennull, so exactly one of the two is ever set and you can consume the result as typed data instead of parsing prose.It's non-streaming only: combining it with
stream=truereturns400, and a schema that can't drive structured output returns422- both before any usage is charged. Conformance is strict onmode=maxand best-effort with automatic retries onfast/smart. -
Steer the agent with a system prompt - a new optional
system_promptshapes the agent's tone, role, or output format (e.g. "answer as a fixed-income analyst, in bullet points"). It layers on top of the built-in instructions and can't override the agent's persona or safety guardrails. -
International equity coverage - the agent can now pull quotes, fundamentals, financials, and technicals for stocks listed in Indonesia, Brazil, and Vietnam, and resolve a company name to its ticker. The existing US-equity tools remain.
-
Streaming restored on
fastandsmart-stream=trueon these modes again streams the answer chunk-by-chunk instead of returning it as a single block at the end. -
Lower latency on repeat and social-data requests - multi-turn requests that resend
message_historykeep more of the response cache warm onfast/smart, and requests that read social-sentiment data no longer stall when that upstream is briefly degraded.
v0.1.3
2026-05-26
- Failed streaming runs are no longer billed -
POST /agent/askwithstream=truepreviously charged you for runs that errored after the response started, whilestream=falsewas already free on the same failure. Both modes now refund, so you only pay for runs that finish cleanly. mode=maxrepairs tool-call pairs in long conversations - sending amessage_historythat contained a tool call without its paired result could trip an upstream400and abort the run. The agent now repairs the pairing before each step, so a multi-turnmaxsession no longer fails late on that error.- Trade setups with a cost check - when your prompt expresses any directional view on a crypto, stock, or commodity, the response now includes a structured trading setup (entry, stop, take-profits, R:R) alongside the prose. Setups whose allocation-weighted R:R doesn't clear an assumed fee + slippage cost are rejected outright instead of being returned as marginal trades.
v0.1.2
2026-05-23
- Usage alerts - you now receive an email the first time your request
usage reaches 80% of your plan's quota for the billing period. It shows
how many requests you've used, how many remain, and when the quota resets,
so you can act before requests start returning
429. Manage notifications from your dashboard.
v0.1.1
2026-05-16
- Request size limits on
POST /agent/ask- requests are now validated against published limits:promptup to 32,000 characters, andmessage_historyup to 50 messages with each message'scontentup to 32,000 characters. Oversized requests are rejected with422before any quota is spent. The limits are published in the OpenAPI schema.
v0.1.0
2026-05-15
- Initial public surface:
GET /health,POST /agent/ask.