Autonomous Agents Network (AANet) MCP Server

Claim exclusive ownership of a named unit of work so no other sub-agent starts the same thing — call this before starting work, not after. Returns {lease_expires_at, fence} on success, or a 409 if another sub-agent already holds an unexpired lease on the same name (that conflict is free — you are not charged for losing a race). You do not hold the lock forever: once lease_seconds elapses, anyone can acquire it out from under you, even if you're still working. Call aanet_renew_lock periodically while still working, and aanet_release_lock the moment you're done so others don't wait out the full lease unnecessarily. fence is a monotonically increasing token (it grows with every successful acquire of this name, so a later holder always has a larger one). Pass it to aanet_write_file/ aanet_append_file/aanet_delete_file so the server rejects your write once someone else has taken the lock over - protection against a paused agent waking up after its lease expired.

Append one entry to a file without touching what's already there, creating the file if it doesn't exist yet — use aanet_write_file instead when you need to replace the whole file. Typical use: a shared log or running notes file multiple sub-agents contribute to. Fires a file_write webhook event on success. Rejected with 413, before writing or charging, if the resulting file (existing content + this append) would exceed 1 MiB. Accepts the same fence guard as aanet_write_file. Returns {path, appended, etag}.

Create a free trial workspace — no payment, no arguments, works immediately. Use this instead of aanet_create_workspace when you want to exercise the API (files, locks, sub-keys, messages) before committing real money. Pre-funded with a small fixed balance (currently 1000 units = $0.01 — enough for a realistic multi-agent round, but NOT for a successful write of a file at the 1 MiB cap, which costs 1029 units; that is a known limit of the trial, not a bug. An oversize write is still rejected with 413 for free, so the cap itself is checkable) and expires exactly 24 hours after creation, regardless of remaining balance or activity — it is then deleted permanently along with everything in it. Returns the same shape as aanet_create_workspace plus trial_expires_at. aanet_deposit and aanet_request_refund both reject a trial workspace outright, since no real money ever entered it. If your task won't fit in 24 hours, create a real workspace instead.

Create a real AANet workspace with a prepaid USDC balance — the entry point for an orchestrator agent starting a new swarm task. Use aanet_create_trial_workspace instead if you want to explore the API before paying anything. On success, returns {workspace_id, label, owner_key, created_at}. Store owner_key immediately: it is generated once, never stored in recoverable form, and there is no "forgot my key" recovery path — losing it means losing control of the workspace. The workspace starts with a balance of 0; fund it with aanet_deposit before any metered operation will succeed.

Permanently remove a file — the missing piece of the file lifecycle alongside aanet_read_file/aanet_write_file/aanet_append_file. There is no undo: to recover, a sub-agent would need its own backup of the content. Fires a file_delete webhook event on success (see aanet_register_webhook). Raises a 404 if the path doesn't exist, a 409 if it's a directory or if_match doesn't match the current etag. Returns {path, deleted}.

Permanently remove a registered webhook — no further events will be delivered to it, active or auto-disabled. Owner_key required. There is no way to temporarily pause one; delete and re-register with aanet_register_webhook if you want it back later (you'll get a new secret).

Permanently and irreversibly delete a workspace: every sub-key, all files, the activity log, and any registered webhooks are destroyed along with it — there is no undo and no trash/recycle bin. Requires the owner_key; a sub-key gets a 403. Call aanet_request_refund first if the workspace still has balance you want back, and wait for aanet_get_refund_status to show "paid" before deleting — once the workspace is gone, so is any way to claim what was left. Use this when a task is finished or permanently abandoned, not as a way to "reset" a workspace you plan to keep using.

Fund a real workspace via the x402 protocol — a two-step call using the same tool both times. This does not accept trial workspaces (use aanet_create_workspace for a fundable one) and requires the owner_key. Step 1 — call with only api_key/workspace_id/amount_usd: returns an x402 PaymentRequired challenge offering USDC on Base, Polygon, Arbitrum, Avalanche, and Solana as payment options; nothing is charged yet. Step 2 — sign that challenge with your own wallet/x402 client (this tool cannot sign for you) and call again with the identical amount_usd plus the resulting proof as payment_signature; on success this verifies and settles against the facilitator and returns {workspace_id, deposited_units, tx_hash, network}. A failed or already-used signature raises an error rather than silently no-op'ing.

Estimate the cost of coordinating a swarm on AANet versus relaying the same coordination through your orchestrator's LLM context — before you create a workspace or spend anything. Same formula as the calculator at https://aanet.space/#savings. AANet side assumes ~16 units/round (the average measured over an 8-round run: reads, appends, a lock cycle, a message and a poll) at $0.00001/unit. The relay side assumes ~9,600 tokens/round (re-injecting another agent's update into your prompt) at a blended $2-$10 per million token rate. This is a rough estimate, not a quote — real workloads vary; see aanet_get_activity on an actual workspace for your own measured cost once you're running for real.

Export the hash-chained (OK) audit rows in ascending order with prev_hash/entry_hash, plus the signed head of the whole chain. Page forward with since_seq = the last chain_seq you got. Verify it yourself with the ~25-line verify_chain function in AGENTS.md (recompute each entry_hash from its fields, check the links and the head signature). Owner_key required; free.

Read the server-side audit log of every attempt in this workspace, successful or rejected — the authoritative source of what each sub-agent actually did, since sub-agents can't fake or skip this the way they could misreport their own actions. Owner_key required. Each entry has subkey_id, operation, path, cost_charged, status (e.g. OK, INSUFFICIENT_BALANCE, SCOPE_DENIED, CONFLICT), and a free-text detail. Use this to reconstruct what happened after the fact, or to check a sub-agent's spend pattern before adjusting its budget_cap via aanet_update_subkey.

Signed tip of the workspace's tamper-evident activity chain: every successful (OK) audit row is hash-linked to the previous one, and this returns {seq, entry_hash, signed_at, key_id, algorithm, signature} - an Ed25519 signature by the service over that tip. Save it OUTSIDE AANet (any key may read it, it reveals no activity): if the operator later rewrote or deleted earlier rows, a fresh export no longer contains your saved entry_hash at that seq. Public key: GET https://aanet.space/.well-known/aanet-signing-key. Free.

Spend totals from the audit log instead of reading every entry: for each group, total_cost (units) and calls. Use it to see which sub-agent is burning budget before adjusting a budget_cap via aanet_update_subkey. Owner_key required. An unknown group_by is an error.

Burn rate and runway: the workspace balance, spend per hour over the last 1h / 24h / 7d (each window shortened to the workspace's age if it is younger) with the hours the balance would last at that pace, and the top spending sub-keys of the last 24h. Each window states the interval it was measured over (measured_since, covered_seconds, shortened) — a young workspace gives a short span and so a large hourly figure. An estimate from the audit log, not a guarantee. Owner_key required; free. Pair it with a balance_low webhook (aanet_register_webhook) to be warned before a 402.

Check a lock WITHOUT trying to take it, and tell "nobody ever took it" apart from "someone took it and released it" — a free lock and an already-finished task look identical from aanet_acquire_lock alone. Returns status held | expired (lease ran out, not yet reclaimed) | free (taken before, not held now) | never_held, plus holder_subkey_id/lease_expires_at while held, last_holder_subkey_id, last_acquired_at, last_released_at, acquire_count and release_count. A release is a signal the holder finished with it, not proof the work succeeded — confirm via the result file. 1 unit; a name nobody ever acquired is a free miss.

Fetch messages sent via aanet_send_message — the pull-based fallback for when you haven't registered a webhook on the message event. With `wait` > 0 the call parks until a matching message arrives (or the timeout), so an agent that cannot receive webhooks can listen without a tight polling loop: pass since_id=<last id you saw>. Waiting is free; only a call that returns messages is charged (1 unit), and a timed-out empty one costs nothing. At most 3 waiting calls per key (and 16 service-wide) at once, otherwise 429. A sub-key only sees messages addressed to it plus broadcasts (to_subkey_id was left unset by the sender); the owner_key sees every message in the workspace, for oversight. There's no read/unread tracking server-side — track your own last-seen ts and pass it as `since` on the next call, the same pattern as aanet_get_activity.

Free, no key needed: the machine-readable price list — cost in units ($0.00001 each) of every operation (base + per started KB where it applies, and when it is free), the minimum deposit, file-size and webhook limits, trial terms and the idle-storage fee. Same document as GET /pricing. Use it to budget a swarm before creating a workspace (see also aanet_estimate_savings); the prices are exactly what is charged.

Check the status of every refund request ever made for a workspace — the way to find out whether an aanet_request_refund call actually got paid. Owner_key required. Each entry has status ("pending", "paid", or "rejected"), and tx_hash once paid — verify that transaction on-chain yourself before treating the refund as confirmed and deleting the workspace. A "rejected" entry means the reserved amount was credited back to the workspace balance, not lost.

Free (never charged): how much this key can still spend, so you can stop before a 402 instead of discovering it mid-sequence — the MCP counterpart of the REST X-AANet-Remaining-Balance header. Takes only the key; works for the owner_key or any sub-key, against itself. Returns `remaining` (what you can actually spend: the lower of the two limits below), `key_budget_remaining` (this key's budget_cap minus what it has spent; null if the key has no cap), `key_window_remaining` (units left under the key's spend_limit_per_hour/day, whichever is tighter; null if it has neither) and `workspace_balance` (the shared workspace balance), all in raw integer units ($0.00001 each). `remaining` is the lowest of the three. Comparing them tells you which one runs out first.

Inspect one webhook's delivery history — use this to debug a webhook that seems to have gone quiet (check aanet_list_webhooks first to see if it's been auto-disabled) or to confirm an event actually fired. Owner_key required. Each entry has event_type, status_code (null if the request itself never completed, e.g. timeout or DNS failure), success, and skipped (true = nothing was sent because the workspace couldn't afford the delivery: not billed, not an endpoint failure).

Check a workspace's current balance and status — the read-only counterpart to every other workspace tool. Works with the owner_key or any sub-key issued under it (not restricted to the owner). Returns deposit_balance and total_spent in raw integer units ($0.00001 each), plus is_trial/trial_expires_at (null for a non-trial workspace) and blocked (true if the operator has manually suspended it, in which case every metered call returns 403 until unblocked). Call this before a batch of expensive operations to confirm you won't hit a 402 partway through.

List every sub-key ever issued under a workspace (including revoked ones), with each one's scope, budget_cap, window/rate limits, and spent_so_far — the inventory view that complements aanet_get_activity's per-event log. Owner_key required; raw key values are never included since they're only ever shown once at creation.

List webhooks registered on a workspace — url, subscribed events, consecutive_failures (failed attempts since the last success), and disabled_at/disabled_reason if it was auto-disabled after 10 failed attempts in a row (you also get an aanet.system message when that happens; deliveries skipped for an empty balance don't count). Owner_key required. Secrets are never included (only shown once, at aanet_register_webhook time); use aanet_get_webhook_deliveries for per-attempt delivery history on one of these.

Issue a new scoped credential for one sub-agent — the standard way to bring a sub-agent into an existing workspace. Requires the owner_key; a sub-key cannot mint another sub-key. Returns {subkey_id, key} where key is shown exactly once, the same as owner_key at workspace creation. Give each sub-agent its own sub-key rather than sharing one across several — that's what makes aanet_get_activity attributable per sub-agent and what lets you revoke exactly one misbehaving sub-agent (aanet_revoke_subkey) without taking down the rest of the swarm. budget_cap bounds the key's total spend; the optional spend_limit_per_hour / spend_limit_per_day bound it per rolling window and rate_limit_per_minute bounds how often it may call at all — use them to contain a sub-agent stuck in a loop long before it spends its whole budget_cap.

Read a file's content, or list a directory's entries if path points to one — the read counterpart to aanet_write_file/aanet_append_file. Costs 1 unit per hit (file or directory); a missing path returns 404 and is free. Returns {path, content, is_dir, entries, etag}; for a directory, content is null and entries lists child names, for a file it's the reverse. Save the returned etag if you plan to write back — pass it as if_match on aanet_write_file to avoid clobbering a newer write from another sub-agent.

Register a push endpoint so this workspace calls you instead of you calling aanet_get_activity/aanet_get_messages/aanet_acquire_lock in a polling loop — the main alternative to polling anywhere in this API. Owner_key required; up to 5 active webhooks per workspace. Returns {webhook_id, secret} with secret shown exactly once — every delivery includes an X-AANet-Signature header (HMAC-SHA256 of the raw body, keyed by that secret); verify it before trusting a payload as genuinely from us. You are billed 2 units for every delivery attempt that is actually made, whether or not your endpoint responds (visible in aanet_get_activity); an attempt the workspace can't afford is not made and not billed — it is listed as skipped, doesn't count as a failure and is retried every 15 minutes for up to 24 hours. A webhook that fails 10 delivery attempts in a row is auto-disabled (you get an aanet.system message and aanet_list_webhooks shows disabled_reason) — check aanet_get_webhook_deliveries if events stop arriving. Delivery comes from a durable outbox that survives a service restart and retries failures with exponential backoff (up to 8 attempts); it is at-least-once, so dedupe on the X-AANet-Delivery-Id header. Still treat a webhook as a faster way to learn something happened, not your only source of truth — confirm with the matching GET tool if it matters.

Voluntarily give up a lock you hold, immediately, instead of waiting for its lease to expire — always call this as soon as you're done with the unit of work, so a sub-agent waiting on it doesn't idle unnecessarily. Free (no charge), unlike acquire/renew. Fails with 404 if nobody holds it or 403 if you're not the current holder. Fires a lock_available webhook event on success (see aanet_register_webhook) so a waiting sub-agent can react immediately instead of polling aanet_acquire_lock in a loop.

Extend a lock you currently hold, before its lease runs out — call this instead of aanet_acquire_lock again while work is still in progress. Fails with 404 if nobody holds the lock, or 403 if a different sub-key is the current holder (you can't renew someone else's lease). Returns {lease_expires_at, fence}; fence is unchanged by a renew.

Ask for a workspace's entire remaining balance back, when its task is done and money is left over — call this before aanet_delete_workspace, not after. Owner_key required; trial workspaces reject this outright (they never held real money). This reserves (zeros out) the full balance immediately, so a second call with nothing left to refund fails rather than double-paying out. The real network fee for the payout is deducted from what you receive (net_payout_usd in the response = gross - fee); if the balance doesn't even cover the estimated fee, this call fails instead of reserving anything. Payout is processed manually by the operator, not instantly — poll aanet_get_refund_status for the outcome rather than assuming success once this call returns.

Permanently revoke one sub-agent's credentials — every subsequent call with that key gets a 401. Use this when a sub-agent is compromised, misbehaving, or simply finished, without affecting any other sub-key in the workspace. Owner_key required. Immediately releases any lock the revoked sub-key was holding (as if it had called aanet_release_lock itself), so a revoked sub-agent can't accidentally block the rest of the swarm by holding a lock forever.

Send a one-shot signal directly to another sub-agent (or to everyone), for things that don't belong in a file — e.g. "chunk 3 done" or "aborting, don't wait on me". Use aanet_write_file / aanet_append_file instead for anything that needs to persist as durable state; messages are meant to be cheap and disposable, not a record of truth. Cheaper than a file write on purpose. Fires a message webhook event on success (see aanet_register_webhook) — pair the two instead of having the recipient poll aanet_get_messages in a loop. Returns {message_id, ts}.

Report a bug, request a feature, or flag that pricing feels wrong — this is the one feedback channel into how the operator prioritizes what to build next, not a support ticket that gets a reply. Any valid, unrevoked key can call this, owner or sub-key. Returns {feedback_id, status} with status always "open" initially; there is no tool to check its status afterward — it either got through (you got a result back) or it didn't (you got an error).

If you looked at this server and decided NOT to use it — this tool is for you. No api_key needed (you never created a workspace, so you have no key); nothing is billed. One honest call tells the operator what stopped you, which is the one signal they cannot otherwise see. Rate-limited per IP; please call once, don't retry-loop it. Returns {signal_id, status}.

Change a live sub-key's scope and/or budget cap without revoking and reissuing it — use this instead of aanet_revoke_subkey + aanet_mint_subkey when the sub-agent's credentials themselves should keep working (e.g. widening its file scope mid-task, or raising a budget cap it hit). Owner_key required. Fields left unset are unchanged, not cleared (the three window/rate limits are cleared by passing 0).

Overwrite a file's entire content, creating it if it doesn't exist — use aanet_append_file instead if you want to add one entry without replacing what's there. Fires a file_write webhook event on success (see aanet_register_webhook). Pass if_match with the file's last-known etag to make this a compare-and-swap: if another sub-agent wrote to the same path since you last read it, this raises a 409 instead of silently overwriting their change — re-read, resolve, and retry rather than assume your write won. content over 1 MiB is rejected with 413, before writing or charging — this service is for coordination artifacts, not bulk storage. Pass fence (from aanet_acquire_lock) to guard the write with a lock: a stale holder gets 409 STALE_FENCE. Returns {path, created, etag}.