Documentation

Pact docs

One guide for both surfaces: Pact for everyday work, and Pact Onchain for public blockchain data. How to give each a job, and what the preview can do today.

Pact

Introduction

Pact is an autonomous work agent. You give it a job and any relevant context. Pact understands the context, works through the task, and returns a useful result.

It is not a chat window. You do not steer it message by message. You describe what you need done, and you get the finished piece of work back: a structured document with a summary, findings, risks, a recommendation where one is called for, and the sources it drew on.

Preview

This release is a preview. Runs are real, with real limits and real tool calls; how much they can do depends on which providers are configured. The preview status section lists exactly what works.

The four stages

Every job moves through the same four stages. This is the short version; the How it works page is the long one.

  1. Input

    You describe the job and attach what it needs: files, links, a repository, data, or just an idea.

  2. Understand

    Pact reads the context and works out the goal, the constraints, and what is missing.

  3. Work

    It researches, analyzes, and works through the task, checking the result against the job.

  4. Result

    You receive clear, structured output with its sources, ready to use.

See each stage with an example job

Getting started

  1. Open the app. It starts on one question: what do you need done?
  2. Write the job the way you would brief a capable colleague. Say what you want back.
  3. Add context with Add files, Add links, or GitHub. More relevant context gives a better result.
  4. Press Run, or press Ctrl+Enter in the input. Pact works through the stages and opens the result when it is done.
  5. Find past jobs and their results under Runs. Copy a result or download it as Markdown.

Writing a job

A good job states the outcome, the criteria, and the scope. Pact can start from less, but it then has to guess.

Too thin

“Suppliers?”

Better

“Compare these four suppliers on price, lead time and risk, and tell me which one to use for the first order.”

  • Name the outcome. A decision, a summary, a plan, an explanation. Say which.
  • Name the criteria. If price matters more than speed, say so.
  • Name the scope. This order, this quarter, this part of the codebase.
  • Say what to return. A table, a one-page brief, a ranked list. Pact will pick a sensible shape if you do not.

Context

Context is everything you hand over with the job. Three kinds can be added from the composer.

Files
Documents, spreadsheets, exports, PDFs. Pick as many as the job needs. Each file is uploaded, checked, and read on the server; tables get their statistics computed in code.
Links
Paste a URL and press Enter or Add. The host and path are shown as a chip. The page is fetched and read when a reader provider is configured; otherwise the step is reported as unavailable.
GitHub
Enter a repository as owner/repository, or paste its GitHub URL. With a GitHub token configured, Pact reads the overview, key files and file tree through the API.

Every piece of context you add is listed under Library, next to the run that used it, and appears first in that result's sources.

Runs and results

While a job is running, the app shows which stage it is in.

Understanding
Reading the job and every attachment; working out the goal and what a good result looks like.
Analyzing
Pulling out figures, claims and open questions; deciding what the result needs.
Researching
Filling the gaps the context does not cover; cross-checking what was found.
Working
Drafting the result and checking it against the job.
Done
The result opens and is kept under Runs.

A finished result has a fixed shape, so you always know where to look.

  • Executive summary. The whole answer in one paragraph.
  • Highlights. Two or three figures that carry the result.
  • Findings and observations. What was found, point by point.
  • Important risks. What could make the result wrong or the decision go badly.
  • Recommendation. What to do, when the job calls for a decision.
  • Sources. Your attachments first, then everything else the result drew on.

Use Copy to put the result on the clipboard as Markdown, or Download .md to save it as a file. Both include every section and the sources.

The app

Home
The composer, three suggested jobs, and your five most recent runs.
Runs
Every run, newest first, with a search box that matches the job, the title and the result.
Library
All context you have given Pact, filterable by files, links and repositories.
Integrations
What is available in the composer today, and what is planned.

On small screens the sidebar becomes a top bar and a tab bar. New run always returns you to the composer.

Examples

These are the kind of tasks Pact is built for. Each opens in the app with the job pre-filled.

More example jobs

Pact Onchain

Overview

Pact Onchain is the same product, pointed at public blockchain data. It lives at onchain.runpact.family and works the way everything else in Pact works: give it a job, get finished work back, with every figure traced to its source.

It covers four chains: Robinhood Chain, Base, BNB Chain and Solana. You can ask about a token, a wallet, a transaction, or a chain as a whole.

Read only

Pact Onchain never holds keys, never asks you to connect a wallet, and never sends a transaction. Nothing it returns is investment advice: it reports what the chain and the public data show, and leaves the decision to you.

Open Pact Onchain

Getting started

  1. Open Pact Onchain. It starts on one question: what do you want to know?
  2. Pick a chain. The four cards above the composer show each chain and how much of it is covered.
  3. Write the job. Name the token by symbol ($BRETT) or paste its contract address; paste a wallet address or a transaction hash as it is.
  4. Press Run. Pact plans the checks, reads the chain and the market data, and opens the result when it is done.
  5. Find past jobs under Runs. Each row carries the mark of the chain it ran on.

Onchain runs are kept apart from your other Pact runs: each surface lists only its own, and the chain tools are only available here.

Chains and coverage

Coverage is what Pact can read for a chain today. The composer shows the same status next to each chain, and a result says so whenever something it needed was not available.

  • Robinhood ChainPartial coverage

    Tokens, tokenized stocks, trending and new tokens, screens, candles, large trades, wallets, transactions and DeFi totals. Market data comes from a single indexer, and there is no honeypot or tax check: risk screens here cover contract verification, holder concentration and scam flags.

  • BaseFull coverage

    Everything in this guide, including honeypot and tax checks, liquidity locks, and the token approvals a wallet has granted with the risk of each spender.

  • BNB ChainFull coverage

    Everything in this guide, including honeypot and tax checks, liquidity locks, and the token approvals a wallet has granted with the risk of each spender.

  • SolanaFull coverage

    Everything in this guide except token approvals, which Solana does not have. Adds mint and freeze authority, a verification and organic-activity score, priced wallet holdings, and launchpad stages such as Pump.fun graduation.

If the job names a chain, that chain wins over the one picked in the composer. If it names none, the picked chain is used.

What you can ask

One job can combine several of these. Ask for due diligence on a token and Pact runs the identity, security and volume checks together; ask what moved over a month and it adds the price history. Each example opens in Pact Onchain with the job pre-filled.

Writing an onchain job

The same rules as any Pact job, with a few that are specific to chains.

  • Name the token precisely. A symbol works ($JUP), but symbols are not unique. When several tokens share one, Pact takes the one with the most real trading volume. Paste the contract address when it matters.
  • Say the period. “Over the last 30 days”, “this week”, “today”. A period is what brings in price history, and it decides between daily, hourly and minute candles.
  • Give the thresholds. “Trades above $10k”, “at least $20k liquidity”, “under 3 days old”, “top-ten share under 40%”. Without them Pact uses its defaults and states them in the result.
  • Compare in one job. Name up to six tokens and ask for a comparison; you get one table instead of six separate answers.
  • Say what the wallet question is. “What does it hold” and “is it safe to interact with” are different checks. Ask for both if you want both.

Too thin

“BRETT good?”

Better

“Is $BRETT on Base safe to trade? Check the contract, liquidity locks and holder concentration, and show what the price did over the last 30 days.”

Reading the result

An onchain result has the same fixed shape as any Pact result: summary, highlights, findings, risks and sources. A few things are worth knowing when you read one.

Security screens
A screen is a set of automated checks, not an audit. A clean one is reported as “no flags found by the checks run”, never as “safe”. Flags are listed with what they mean.
Sources
Figures carry a reference to the source they came from, so you can see which reading a number rests on and check it yourself.
Single-source figures
When a figure comes from only one source, the result says so. That is the normal case on Robinhood Chain.
Missing data
If something could not be read, the result names it instead of estimating. A gap is listed as a limitation, not filled in.
Snapshots
Prices, volume and holders are as of the moment the run read them. Fast markets move; run the job again for a fresh reading.

Limits

  • Read only. No wallet connection, no signing, no trading, no transfers.
  • No advice. No buy or sell calls and no price predictions. Pact describes liquidity, volume, holders, authorities and activity.
  • Four chains. Robinhood Chain, Base, BNB Chain and Solana. Other chains are not read.
  • Public data only. Pact sees what the chain and public market data show. It does not see private order flow, off-chain agreements or who is behind an address unless that is public.
  • Recent windows. Large trades and volume tracking cover the last 24 hours. Price history goes back up to a year in daily candles.
  • Size of a result. Comparisons take up to six tokens, screens return up to 25 rows, and wallet activity lists up to 25 recent transactions.

Reference

Preview status

Being clear about this matters more than looking finished. It covers both Pact and Pact Onchain.

Works today

  • Describe a job, add files, links and a GitHub repository, and run it.
  • A real run: a plan, tool calls with hard limits, progress as it happens, and a structured result.
  • Files are uploaded and read on the server: PDF, DOCX, XLSX, CSV, JSON, TXT and MD, with table statistics computed in code.
  • Web search, page reading and GitHub run when their providers are configured; otherwise they are reported as unavailable, not faked.
  • Find past runs under Runs, search them, and reopen any result. Copy a result or download it as Markdown.
  • Cancel a run while it is in progress.
  • Pact Onchain on four chains: token analysis, security screens, price history, volume, large trades, comparisons, new tokens, screens, DeFi, wallets and transactions.

Not yet

  • Running without a language model. If no configured model responds, the run fails with a clear error rather than producing a substitute result.
  • Accounts. Runs and files are scoped to a per-browser id, not a login.
  • Onchain actions. Pact Onchain reads public data; it does not connect to a wallet, sign, trade or send anything.
  • Chains beyond the four listed, and honeypot or tax checks on Robinhood Chain.
  • Integrations beyond the composer. Cloud storage, team workspaces and databases are listed as planned, not available.

When providers, accounts or a hosted database change what Pact does with your data, this page, the privacy notice and the terms are updated first.

FAQ

Where does my data go?
To the Pact server. The job, the files you attach, the progress and the result are stored there so you can come back to them. Links and repositories are only fetched when a provider is configured. See the privacy notice for the full list.
Which model works on my job?
Auto picks by the job: light tasks go to a fast model, heavy research, code and multi-source analysis go to Claude, with Gemini and open-weight models as fallbacks. You can also pick a model yourself in the composer. The run page and the result show which model actually did the work.
Why was a step skipped?
Because the tool it needed has no provider configured, or the run reached its step limit. Skipped steps are listed in the result’s Limitations section and in the run’s steps.
Can I delete a run?
Not from the interface yet. Runs are stored on the server; deletion is planned alongside accounts.
Do I need to connect a wallet to use Pact Onchain?
No. It reads public data only. It never asks for a wallet connection, a seed phrase or a signature, and it cannot move funds.
Is a clean security screen a guarantee?
No. It means the checks that ran found no flags. Contracts can change, checks can miss things, and a token with a clean screen can still lose its value. Nothing Pact Onchain returns is investment advice.
Why does Robinhood Chain say partial coverage?
Its market data comes from one indexer, so figures there cannot be cross-checked against a second source, and honeypot and tax checks are not available for it yet. Results say so where it matters.
Does it work on a phone?
Yes. The app has a top bar and a tab bar on small screens, and results are laid out for a narrow column.