> ## Documentation Index > Fetch the complete documentation index at: https://docs.traderframe.io/llms.txt > Use this file to discover all available pages before exploring further. # Trade Agent > An AI agent on your bot that finds or judges entries under your caps — proposing them, placing them within limits, or running its own exits, at the tier you choose ## What it is The **Trade Agent** is a signal source you add to a bot, like TradingView or Telegram. On a schedule it looks at the bot's pair, or it judges the signals another source sends, and it decides an entry with its reason and levels. What happens next depends on the tier you set: at **Suggest** nothing is placed until you tap **Place it**; at **Enter** it places within your caps and tells you after; at **Manage** it owns the trades it opens: size, exits, trail and the ladder. Every look it takes is written down, so you can always see what it saw, what it decided, and why. The agent is judged the way a trader is: by its record. It runs inside the same rails as every other source — Risk Management, the news filter, your daily limit and your plan's limit all apply to anything it places. ## Adding it to a bot 1. In the builder, press **+** under **Signal sources** and choose **Trade Agent**. 2. Set it up in the panel (below) and press **Save**. Closing the panel with **Done** also saves any changes. 3. **Activate** the bot. A sheet tells you which account it runs on, when the first look is, and what it can spend. Only the bot's own pair is tradable. Another pair means another bot. ## The panel * **Autonomy** — three tiers, one ladder. **Suggest**: the agent proposes, you approve each trade. **Enter**: it places within your caps without a tap and you are told after each entry; a gate can still refuse it. **Manage**: everything Enter does, and it owns the trades it opens. It sets the size from your martingale ladder — the multiplier on the step the bot stands on after a loss, back to the base after a win, never above your cap — puts its own stop and target on the order whatever the execution, may set its own trail in R that replaces the Exit rule's ladder for that trade, and manages the position on every look: tighten the stop or exit, never widen, never add. Its one goal is the most sound trades and a growing balance. Manage runs on MT5 and crypto accounts; on binary options it records what it would have done. Prove any tier on a demo account before a live one. * **Style** — which playbooks it may take. **Trend and range edges** (the default) allows all six. **Trend only** takes setups with the context bias and no range-edge fades. **Scalp** runs the same playbooks every five minutes on 5m and 15m candles with tighter stops and targets. * **Entries** — **Judge a signal source** holds every signal from the bot's other sources and comes back with a proposal or a skip. **Find its own** looks at the pair on a clock. * **Look every** — 15 minutes to 4 hours, on the candle close. * **Only during** — the market sessions it may look in, shown in your own time. Outside them it sleeps and spends nothing. * **It reads** — the timeframes it uses. The first is the one it enters on. What each look contains is listed under [What it sees](#what-it-sees). * **Hard limits** — max size, trades per day, daily loss stop, cooldown after a loss, reward-to-risk floor, looks per day. The engine enforces these; the model cannot argue with them. * **Model** — the deciding model, and what a look costs: **Sonnet** is the default at 1 agent credit a look; **Opus** is the biggest model at 5 credits a look. * **Tell me about** — which events you are told about, and a **Telegram** row: link the TraderFrame bot once and proposals arrive on your phone with **Place it** / **Skip**, stops and daily reports as a line. Answering there is the same as answering in the app; a tap in both places cannot place twice. * **Look now** — one look outside the schedule, so you can watch it think. ## What it sees Every look is one snapshot, built from public data and the bot's own record. Nothing in it comes from your exchange account. * **Candles** — the last 100 closed bars on the entry timeframe and 60 on each higher one: open, high, low, close, in UTC. The bar still forming is left out. * **Read from the candles** — ATR(14), session high and low and where price sits between them, structure (higher lows, lower highs, range, mixed), the last two swings, bars since the session extreme. * **Volume** — on crypto, the last bar and the last five bars as a share of the typical bar (the median of the 20 before). Forex and Deriv feeds carry no volume. * **The clock** — the session it is in, and the next red news release and how far away it is. * **The bot** — open position or not, trades today against the cap, ladder step, day P\&L, last loss. Plus your caps. * **Its memory** — its own last 20 decisions on this bot with their outcomes. On crypto it also reads **the market around the candles**, from the exchange's public feed at the moment of the look: * **Spread and costs** — bid and ask, as a price and as a share of price, with 24-hour turnover, and the round-trip cost (spread plus fees on crypto). The engine refuses a stop closer to the entry than three spreads and a target inside three round-trip costs; on a scalp those are the checks that matter. * **Funding** — on perpetuals, the current rate and which side pays. Beyond ±0.05% per interval the paying side is crowded: an entry on that side is a skip unless it is a pullback in a higher-timeframe trend. * **Depth** — the resting orders within half a percent of price, folded into a bid share and the nearest large wall on each side. A target is placed before a wall, never past it. The book alone is never a reason to enter. It does not see the raw order book, open interest, liquidations or your wallet. ## What it reads on every close Before any model is called, the engine reads the chart the way a discretionary day trader would, and every number goes into the look: * **Levels** — support and resistance from clusters of swing highs and lows, with how many times each held. **Liquidity** — equal highs or lows where stops rest, and **sweeps**: a wick through a level that closed back inside. * **Gaps** — fair value gaps (three-candle imbalances not yet filled) and opening gaps away from the prior close. * **The candle** — pin bar, engulfing, inside bar, strong close. **The regime** — EMA20/50 alignment, volatility against its own average, the last five swings as HH/HL/LH/LL, the bias, the last break of structure and any change of character. * **Context** — a wider timeframe behind your two (4h behind a 5m/1h bot) for the bias and the levels that matter, plus the prior day's high, low and close. From those it finds **six playbooks**, each with a stop beyond the structure that would prove it wrong and a target at the next level: | Playbook | What it is | | ---------------- | ---------------------------------------------------------------------------------------- | | Trend pullback | Trend on the higher timeframe; price came back to a level, the EMA20 or a gap and turned | | Breakout retest | Broke a level on a real candle, came back to it and held | | Liquidity sweep | Stops taken beyond a level, candle closed back inside; fade it | | Range edge | No trend; a level that held twice or more was rejected | | Gap continuation | In a trend, dipped into an open fair value gap and closed back out | | Gap fill | An opening gap that has started to close; target the prior close | Each playbook carries a quality score and its evidence in words. A close with no playbook is a **free** engine skip that says what it saw. A found setup goes to the model, which takes the best one when it is sound — quality 0.55 or more, the target at least your reward-to-risk floor, the stop beyond the structure — or names the flaw. "Mid-range" on its own is no longer a reason to skip. On the Decisions tab the playbook shows as a chip on the row, and **Open the look** lists everything that was on the table. **The daily report prices looks the way you are charged.** "The agent's day" states the looks as a count times the look rate on your plan, the per-look price of the pack at your plan's scale (15¢ a look up to 100 looks a month, 5¢ up to 1,000, 3.3¢ up to 3,000, 2¢ beyond). It never shows what the model itself cost. **Quiet rows stay out of the list.** While the bot's trade is open the agent keeps looking on its clock but writes nothing; the open trade is on the Trades tab. A budget sleep ("today's looks are used") appears once, not every five minutes. **Every look can be drawn.** Open a look and the chart at the top shows the candles the agent was given, the levels that have held, the pools of stops, the sweep and the open gaps, with the trade drawn over them: Entry, Stop, each trailing step (where profit locks), and the Take profit. Numbered pins mark the playbook's level, the sweep, the walls on the way, and the prices the reason names, with the text under the chart. On a skip the chart draws the best playbook that was on the table, labelled as not taken. **Your Exit rule sets the exits at Suggest and Enter.** When the bot executes at market, every agent entry carries the stop, target and trail from the bot's own Exit rule, and the agent picks entries that fit them — with a trail on, it prefers moves with the momentum to run through the first step: points on MT5 (gold has two digits, so TP 1000 points is $10.00 and SL 500 points is $5.00), ROI percent of margin on crypto. The agent does not invent its own targets there; it judges whether price can reach the point where your profit locks before your stop, and it never skips because the target is small or far. That point is the first trailing step when your rule has a trail, else the take profit. So a rule of TP 2000, SL 300 and trailing at 450 points on gold is read as 300 of risk against 450 of reach, with the 2000 left as the big move the trail is allowed to catch. A level that has held before and sits on the way is named to the agent as a wall to weigh, not an automatic skip. A bot set to Limit execution uses the agent's own levels instead. To scalp gold for 1000 points with a 500-point stop, set the Exit rule to TP 1000 and SL 500 points, choose the Scalp style, and keep execution at market. The agent also keeps a record of how each playbook has done on your bot over the last 30 days. One that has lost most of its last six or more is benched: discounted and named to the model, not banned. ## When it seems to never trade Most looks end in a skip; that is the design. The strip on the Agent decisions tab shows the week in numbers — looks, candidates, entries — and, when there were looks but no entries, a **Why quiet** line with the latest reason. A run of engine skips reading "No playbook: …" is a tape with no level, no sweep and no rejection candle to trade, not a fault. If you want more setups, make sure **Style** is trend and range edges and that your sessions cover the hours the instrument moves. ## Looks and agent credits A **look** is one model call by your agent. It is paid in **agent credits**, the agent's unit the way a trade is the bot's: **1 credit a look on Sonnet** (the default model), **5 credits a look on Opus**. * The checks the engine settles on its own — outside your sessions, a position already open, the daily cap reached, red news inside 30 minutes — are **not charged**. They are free and still logged, once per streak. * A look that becomes a trade you place counts **its credits and one trade**. * **On a 7-day trial** the agent runs on **10 credits, Sonnet only**, with every tier available; Opus and the plan's allowance open when the trial converts. The panel's meter says "of 10 trial credits" and names the date. * The Trade Agent is on **Basic and up**; the Free plan has none. Paid plans include agent credits a month: **Basic 50 · Standard 150 · Pro 400 · Ultimate 1,000**. On Opus that is 10, 30, 80 and 200 looks. Plan credits reset with the billing period. * **Credit packs** top a plan up: 100, 1,000, 3,000 or 10,000 credits bought outright on the pricing page (**Add Trade Agent credits**). Pack credits never expire and are used only after the month's allowance is gone. They stay on your account if your plan ends or changes; only the plan's monthly credits end with the plan. The **Add credits** button appears wherever the cap does. * A look is refused before it starts when fewer credits remain than the bot's model needs. An Opus bot with 3 credits left sleeps until the next period or a pack; switching it to Sonnet lets it use the rest. * When the period's credits are gone the agent does not stop: its free engine checks continue and it simply stops calling the model until the next period or a pack. You are told once. * Where to see them: the sidebar plan card, **Subscription & billing**, the agent panel (which also shows what a look costs on the chosen model), and the **Agent decisions** tab, where each row says what it charged. ## Proposals When the agent decides to enter, you get a card in your notifications with **Skip** and **Place it**, the reason, and the entry, stop and target. It expires after **5 minutes**; a stale entry on a fast market is a different trade. **Place it** walks the ordinary path, so a gate can still refuse it, and the row will say so. Whatever you do, the agent follows the proposal on paper and records what it would have done, so the record grows even when no money moves. ## When it stops itself Three failed model answers in a row and the agent switches itself off: it tells you once, places nothing, and the bot's other sources keep running. The strip and the **Agent decisions** tab show **Stopped** with the reason and a **Resume** button; saving the panel also resumes it. ## The daily report * **The agent's day**: once your day has ended, each bot with a decision gets one short paragraph — the numbers first, then what the skips had in common. It is a card at the top of the tab and a notification; no advice, no predictions. ## The record * **Agent decisions** tab on the bot: every look, skips included, who decided it (an engine rule, the screening model, or the deciding model), the reason, the levels, and **Open the look** to see exactly what it saw. * **Signals** tab: every look that spent credits, and every proposal once decided, as a row tagged **agent look**, **proposal**, **you skipped**, **expired**, **you placed it**, **the agent placed it** or **refused**, with the reason and the credits it cost. A signal the agent held for judging stays under its own source, tagged **held for the agent**. * **Agent skipped** is its own status and filter. A decision not to trade is not a failed signal. Sizes on the tab, in notifications and in the daily report are in the venue's unit as you read it: lots on forex, contracts on futures, USDT on crypto, USD on binary. Levels are quoted to the pair's decimals. A look that reads "No candles: the exchange is rate-limiting candle requests" was settled free by the engine after two tries 2.5 seconds apart; the next slot tries again. ## Your own agent On Pro and Ultimate you can train agents of your own in **Agents** (the Studio) and run them on any bot. An agent is a profile, not a bot: one agent can run on several bots, and a bot picks which agent it runs in the Agent field at the top of the Trade Agent panel. The stock agent stays the default and the baseline. * **Teach** — rules in your own words. What the engine can enforce it enforces free before a look: trading hours in your day, one side only or only with the wider bias, a wider news window, a loss stop, base size past a ladder step, a stop floor in ATR, playbooks switched off. Every other sentence goes to the model as your instruction. Each rule shows how the agent read it; fix a misreading by editing the sentence. Preferences are weighed, not enforced. * **Show** — grade the looks it took: good call, I'd have skipped, I'd have entered, wrong size, with one line why. You can grade on the bot's Agent decisions tab too. **Write lessons** distils the grades into up to eight lessons with counts; the agent reads them before every look. * **Read** — add a strategy PDF, notes, a chart screenshot, a CSV trade log, or a pasted video transcript (open the video, choose Show transcript under the description, select all, copy, paste). Long text is read in pieces. The file is read once into rows you keep or drop; the file itself is never stored. A row that contradicts one of your rules is flagged and you pick which stands. Word and Excel: export as PDF or CSV first. * **Your own playbooks** — on Teach, write a setup of your own: a name, the side, what the market must show, and where the entry, stop and target go. The model reads it on every look and may enter on it even when the engine's six found nothing; the record names it so its results are tracked. Up to 10 per agent. * **Prove** — the ready score (paper 35, grades 25, agreement 40) tells you what the agent has earned so far: Enter at 60, Manage at 85 plus 20 live entries. It advises, it never decides: you pick the tier on each bot's panel, demo or live, and the panel shows the earned tier beside your choice with a caution when you go above it. **Publish** snapshots a version; bots run the last published one, and edits after that are a draft until you publish again. A personal agent's look costs the same credits as the stock agent's. Training, grading, files, lessons and the brief cost nothing. Pro holds 2 agents, Ultimate 5. ## What it will not do * Place a trade without your tap, at the Suggest tier. * Trade a pair the bot does not have. * Exceed a cap, widen a stop, add to a position, or skip a gate. * Charge credits for a failed model call or a check the engine settled itself. * Keep calling a model that keeps failing: three in a row and it stops and tells you.