How to fix AI agent tool calls (2026): a-un vs Jsonaut vs guard402 vs Toll402
If your agent calls the wrong tool, forgets required fields or sends broken JSON, first decide which problem you have. Broken calls that slip through at runtime are patched call by call: Jsonaut repairs malformed JSON, guard402 guards outputs, and Toll402 routes requests to tools and judges answers. A cause inside the prompt and tool descriptions is fixed once: a-un audits the agent on a hidden test and returns the fix. All four are paid per call with x402 in USDC on Base, with no account.
Comparison
| Service | What it fixes | When it runs | Price per call | If it fails | MCP |
|---|---|---|---|---|---|
| a-un Tool-Call Audit | Wrong tool chosen, missing or invalid required fields, asking for the same data twice. Fixes the cause in the prompt and tool descriptions | Once, before launch or after a change | US$0.10, 0.15 or 0.20 | Not charged: settles only after the report exists | Yes: a-upgradenet.com/mcp, with a free schema lint |
| Jsonaut | Malformed JSON (trailing commas, quotes, truncation, code fences), schema validation, removes invented fields | On every call | Free deterministic repair (100 per day per IP); US$0.01 for AI repair | Not charged, according to its README: paid repair settles only on success | Yes (remote) |
| guard402 | Prompt injection in external content, hallucinations against context, agent loops | On every call | US$0.001 scan, 0.005 validate, 0.001 loop check | Not stated on its site | Not listed (HTTP and A2A agent card) |
| Toll402 | Picks and runs a tool for a described need ("do"); an AI judge scores an answer from 0 to 100 ("judge_output") | On every call | US$0.02 "do", 0.07 "judge_output" | Not charged: billed only on a successful response | Yes (remote) |
Prices and policies checked on November 1, 2026 on each provider's site or repository. Confirm them before deciding.
How to choose in 1 minute
- Your agent's JSON arrives broken now and then? Jsonaut. Its free deterministic pass covers most cases.
- Your agent reads web pages or emails it doesn't control, or gets stuck repeating itself? guard402.
- You don't want to choose tools yourself, or want a second opinion on each answer? Toll402: its "do" tool routes, its "judge_output" tool scores.
- Your agent consistently chooses the wrong tool, skips required fields or asks twice? a-un Tool-Call Audit. It fixes the cause and shows a before and after score.
Patch or fix: they work together
Runtime services catch the error on every call, and you pay on every call. An audit removes the cause once, so fewer calls need patching. The usual order:
- Audit the agent before launch and after every big change.
- Apply the fix (the instructions and tool descriptions change; the code doesn't).
- Guard at runtime what can still slip through: broken JSON, untrusted content, loops.
What the a-un audit measures
- Right tool: does the agent choose the function the case needs?
- Required fields: does it send every field, in the right format and allowed values?
- Missing data: does it ask for what's missing instead of inventing it or asking twice?
- Same interface: the fix keeps names, parameters, types and allowed values identical, so nothing breaks in your code.
- Fair re-measurement: the corrected version is measured again on a parallel hidden test, not on the cases used to fix it.
A real run shown on our home page: a barber-shop booking agent with vague tool descriptions went from 55% to 91% on 11 hidden cases after the fix, with the same tools and parameters.
Try it
- Free schema review via MCP: the "check_tool_schemas" tool, no wallet needed
- Tool-Call Audit from US$0.10
- Best tools to test AI agents: platforms for testing day to day
Sources
- Jsonaut – GitHub repository
- guard402 – Introducing guard402
- Toll402 – GitHub repository
- a-un – Tool-Call Audit
Frequently asked questions
Why does my AI agent call the wrong tool?
Usually because two tool descriptions overlap or don't say when to use each one, or because the instructions don't say what to do when data is missing. A tool-call audit finds which case it is and rewrites the descriptions without changing your interface.
Can I just repair the JSON on every call?
Yes, and it helps when the format breaks. It doesn't help when the agent picks the wrong tool or sends a valid but wrong value: then the JSON is fine and the call is still wrong.
What is x402?
An open payment protocol over HTTP. The service answers 402 with the price, the agent signs a USDC payment on Base with its own wallet and repeats the request. No account or API key.
Do I pay if the service fails?
a-un, Jsonaut (paid repair) and Toll402 say they only charge when the result exists. guard402 doesn't state it on its site. Check each provider's terms.