A1 ratingsPrinceton/Stanford NLPmini-SWE-agent

mini-SWE-agent

Princeton/Stanford NLP · agent · Any industry
C
CGrade
41Score
±795% range
#602Rank of 1134
0%From fleets
FR-CAT-00863Catalog no.

The six pillars

Each bar is the score; the bracket is its 95% range.

Can you see what it costs?

Only through a gateway you run · prices published
58±17
Desk research 60 · counts for 25% · A cost API you can call scores highest; an admin console or cloud billing export next; a card statement or emailed invoice is all some products leave you. Published prices add to it.

Can you see how much it's used?

No usage API found
50±17
Desk research 50 · counts for 15% · A usage API (tokens, seats, credits, conversations) a buyer can read.

Can you tie cost to one agent?

Nothing a buyer can read
10±17
Desk research 0 · counts for 15% · Exact (one key or project per agent) beats by-project, by-workspace or by-user.

Can an agent on it report out?

SDK · ~100-line Python agent; routes models via litellm/OpenRouter/Portkey, so a gateway is the natural place to meter it.
66±17
Desk research 70 · counts for 15% · A remote MCP connector scores highest, then a local MCP server, an SDK, an API or a workflow step; copy-and-paste lowest. Any open standard counts, not just FleetRegistry.

Can you control spend and access?

Spend cap: Hard cap · Per-agent keys: One shared key · Audit log: None · Revoke: None · SSO: None
34±17
Desk research 30 · counts for 20% · Hard spend caps (30%), scoped per-agent keys (25%), an audit log you can read by API (20%), revoking by API (15%), SSO with SCIM (10%).

Does it work on small plans?

Small plans don't get the data
10±17
Desk research 0 · counts for 10% · Whether the plan a person or small business buys gets the cost and usage data at all.

Where the evidence comes from

Independent fleets0%Desk research80%Neutral starting point20%

Once three independent fleets report on a pillar, desk research counts for at most a third of it. FleetRegistry's own fleets never count.

Controls on record

Spend capHard cap
Per-agent keysOne shared key
Audit logNone
RevokeNone
SSONone
PricesPublished

Free open-source research agents with no vendor-side controls. The only built-in guard is a local per-run cost limit (mini-swe-agent cost_limit, 0 = off by default; SWE-agent per_instance_cost_limit default 3.0 and total_cost_limit) plus step limits; keys come from the user's own env.

History

11-12 C 4111-13 C 4111-14 C 4111-15 C 4111-16 C 4111-17 C 4111-18 C 4111-19 C 41
Rated since 2026-10-02. Published grades are sealed daily into the transparency log (N4.4). Grade held since 2026-10-02.

Every fact, with its rung and source

Each fact stands on its own source and its own rung of the evidence ladder (Rule N2.1). Verified Observed Declared Inferred. The weakest fact in a pillar sets its weight.

FactValueRungCheckedSources
How a buyer sees cost
Cost
Only through a gateway you run Declared
high confidence
2026-10-25
1 earlier
[1][2] Correct
Prices published
Cost
Yes Declared
high confidence
2026-10-25
1 earlier
[1][2] Correct
Usage API
Usage
None found Declared
high confidence
2026-10-25
1 earlier
[1][2] Correct
Cost per agent
Per agent
None found Declared
high confidence
2026-10-25
1 earlier
[1][2] Correct
How an agent reports out
Reports out
SDK Declared
high confidence
2026-10-25
1 earlier
[1][2] Correct
Per-agent keys
Controls
One shared key Verified 2026-11-15
2 earlier
[1][2][3] Correct
Audit log
Controls
None Declared
high confidence
2026-10-25
1 earlier
[1][2] Correct
Revoke access
Controls
None Declared
high confidence
2026-10-25
1 earlier
[1][2] Correct
Spend cap
Controls
Hard cap Declared
high confidence
2026-10-25
1 earlier
[1][2] Correct
SSO
Controls
None Declared
high confidence
2026-10-25
1 earlier
[1][2] Correct
Small plans get the data
Small plans
No Declared
high confidence
2026-10-25
1 earlier
[1][2] Correct

Prices · Free (MIT)

Notes · Research agent from the SWE-bench team (v2, PyPI 2.4.6, July 2026); the README calls it the default choice over SWE-agent. Costs are whatever your model provider charges.

Scored with the rubric for Agents and agent platforms (Rule N7.1). Each pillar blends a neutral 50, desk facts weighted by rung, and independent fleets. A1 is 85+ with fleet proof, A 70, B 55, C 40, D 25, E below. Part N of the Rules

Run agents on mini-SWE-agent?

Give each one a permanent number, a named person who answers for it, and a record of what it can touch and what it costs. Free.