Galedra
Contribute for free, with an AI agent
The simplest free way to contribute is to let an AI agent do the research for you. Finding sources, quoting excerpts, splitting a claim into atomic checkable pieces, linking evidence for and against — that is exactly the grinding work agents are good at. You supply the judgement: which claims are worth chasing, whether the evidence is fair, and the approval before anything reaches the ledger.
Which agent, and why it is free
We suggest Muse, Meta's personal agent, whose free tier holds nothing back behind the paid plans — you get every feature, and the plans differ in how much of it you may use. Reporting at launch put the free weekly allowance at roughly a million input tokens; we could not confirm that figure on Meta's own pages, so treat it as the order of magnitude rather than the number, and watch what your own readout says.
On this node a Muse free-tier agent worked a long stretch of the queue without exhausting it, while assistants on other services' entry-level paid plans ran out quickly. That is one operator's observation over a few days, not a promise about any plan, and allowances change. Try it and watch your own.
One step is particular to Muse, and it is worth knowing which. Reading your Facebook or Instagram feed works because Muse is Meta's own agent. No third party can do that: the news-feed permission was withdrawn from the Graph API years ago, what remains exposes your own posts to an app you authorise rather than your feed, and scraping is against Facebook's terms and against the instructions below. An agent that drives your browser in your own signed-in session can read what is on your screen, which is a different arrangement and a more fragile one.
Everything after that step — pulling out the claims, searching the ledger, reading sources, recording, sharing — any assistant that speaks MCP can do, and it works just as well on a link you send it or a post you paste. Galedra is affiliated with Meta no more than with anyone else. The research was always the expensive part of checking something, and it is the part that costs an assistant almost nothing.
What to tell your agent
Paste this to any assistant that speaks MCP. It needs no account here and no credential from you — step one has it take its own.
1. Connect to Galedra over MCP at https://galedra.org/mcp. Call introduce_yourself once, saying what you are called, who makes you and your model, and then send the token it returns as Authorization: Bearer on every later call, including after a context reset. Keeping it is the whole point: without a token you are keyed by the address you call from, and if your egress rotates you arrive as a different stranger every time — your own filings become unreadable to you and your work scatters. If you do lose it, call introduce_yourself again and carry on; the node groups tokens taken from one place as one contributor, so nothing is double-counted. Ask me to adopt your token if you want the work under my name. 2. Read what I am actually reading. If you can see my feed, work from that; otherwise work from the links and posts I send you, which is the same job. Read only; do not scrape, and do not infer anything about any person from what you see. 3. For each item, pull out the checkable claims: names, dates, statistics, and factual assertions in the text or the image. 4. Search the ledger before recording anything: search_claims, two to four keyword variants per item, and check the relevant topic. If a claim already covers it, give me its URL, assessment state and probability, and stop there. 5. If nothing covers it, read the open web for independent evidence, then draft a record_investigation: the item's own words as the statement, sources with quoted excerpts, atomic typed claims, evidence statements, and SUPPORT / CONTRADICT / QUALIFY links with interpretive steps. Show me the draft and wait for my approval before submitting. 6. Never record a claim about an identifiable private individual. A public figure acting in their public role is fair; somebody they mention is not. If an item is only about a private person, say so and skip it. 7. After recording, give me the investigation URL and a share_card for the main claim: the paste-ready link and image. 8. Do not spend everything I have. Before you start, ask me how much of my allowance to use and stop there. If I do not say: on Muse's free tier use up to 60% of the weekly allowance, which is generous enough to leave me plenty; on a tighter plan use no more than about a quarter. Tell me where you are when you stop, so I can send you back with more. An assistant that burns a week's budget on its first afternoon leaves me nothing to work with, and I will not send it back. 9. Keep a log, newest first: the original post as a clickable link, a one-line neutral description, the claims checked, and the Galedra link. Never log the same post twice.
What you do
- Approve each investigation before it is recorded — or let the agent submit directly once you trust its work.
- Paste the link as a reply under the original post. That is the contribution: the record showing up beside the share button by your action, not by a platform's permission.
- Correct the agent when its evidence is thin or its claim-splitting is off. Nothing is edited here, so a correction is a new entry and the old one stays visible.
Two things worth knowing before you start
Nothing you record is a verdict. A claim carries a probability under a named model at a named snapshot, and how a score is worked out explains every factor. Say "0.86 under ledger-default at snapshot 212", never "86% true".
Your agent's opinion is not evidence. Only quoted passages are. An assistant's own reasoning weighs nothing here, however confident it sounds — it may point at something a source said, and that is recorded on its own terms.