@unerr-ai/unerr 0.2.7 → 0.2.9

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/README.md CHANGED
@@ -1,15 +1,18 @@
1
- <p align="center">
2
- <a href="https://www.unerr.dev/"><img src="https://unerr.dev/icon-wordmark.svg" alt="unerr — operational intelligence for your codebase" width="320" /></a>
3
- </p>
1
+ <h1 align="center">
2
+ <a href="https://www.unerr.dev/"><img src="https://unerr.dev/icon-wordmark.svg" alt="unerr" width="320" /></a>
3
+ </h1>
4
4
 
5
5
  <p align="center">
6
- <strong>Your AI agent has read your codebase. It was never briefed on it.</strong>
6
+ <strong>Your AI agent has read your codebase. It still can't safely change it.</strong>
7
7
  </p>
8
8
 
9
9
  <p align="center">
10
- <strong>unerr is operational intelligence for your codebase</strong> — the lived history your team carries in its head:<br/>
11
- what's been tried, what broke, what the team decided. One local runtime, <em>behind</em> every MCP your agent already speaks,<br/>
12
- that hands the agent that judgment the moment it starts working instead of letting it relearn your repo, every session.
10
+ Every tool built to help hands your agent <em>advice it can ignore</em> — a memory it has to remember to check,<br/>
11
+ a graph it has to choose to query, a reviewer that only speaks up after the break is already written.<br/>
12
+ <strong>unerr is the guardrail it can't skip.</strong> The moment your agent edits a function, unerr puts the live call graph<br/>
13
+ and the rule you pinned to that exact function <em>into the edit itself</em> — automatically, not on request — and re-anchors<br/>
14
+ that rule when the code moves, so it never goes quietly stale. The 24 callers and the standard it's about to break are<br/>
15
+ on screen <em>before</em> the function changes. Every time. Whether or not the agent thought to ask.
13
16
  </p>
14
17
 
15
18
  <p align="center">
@@ -31,39 +34,59 @@
31
34
  <sub>Zero configuration. Install, restart your IDE, and the next prompt already knows your repo.</sub>
32
35
  </p>
33
36
 
34
- <p align="center">
35
- <sub>Measured, not estimated: the agent lands on the right code while spending <strong>86–90% fewer tokens</strong> getting there —<br/>
36
- same corpus, same tokenizer, with a fidelity gate that discards any "saving" that lost the answer. <a href="./benchmarks/README.md">See the benchmarks →</a></sub>
37
- </p>
37
+ ---
38
+
39
+ <details>
40
+ <summary><strong>Contents</strong></summary>
41
+
42
+ - [The gap nobody else closes](#the-gap-nobody-else-closes)
43
+ - [The pains this fixes](#the-pains-this-fixes)
44
+ - [What changes when you install it](#what-changes-when-you-install-it)
45
+ - [See it in action](#see-it-in-action)
46
+ - [Quick Start](#quick-start)
47
+ - [Who it's for](#who-its-for)
48
+ - [Why a guardrail has to be one runtime, not five tools](#why-a-guardrail-has-to-be-one-runtime-not-five-tools)
49
+ - [How the runtime works](#how-the-runtime-works)
50
+ - [Fewer tokens, as a side effect](#fewer-tokens-as-a-side-effect)
51
+ - [License](#license)
52
+
53
+ </details>
38
54
 
39
55
  ---
40
56
 
41
- ## The old way is over
57
+ ## The gap nobody else closes
42
58
 
43
- Coding agents now write the code. They've read every line of your repo and not one of them has been briefed on it. They don't know what the team tried here and abandoned, why this function drifted, what broke the last time someone touched it, or which decision is load-bearing. A new engineer gets that briefing on day one. The agent starts cold, every single session.
59
+ On a small or greenfield project the agent holds the whole repo in its head and reading the live code is enough — you don't need us. The wall is the *large, existing, multi-contributor* codebase, and it's the same wall every time: the agent can't fit the whole thing in context, so it acts on the slice it can see and never reads the rest. It changes a signature and breaks 7 of 24 callers it never read. It writes a fourth copy of a pattern your team standardized months ago even with the rule spelled out in `.cursorrules`. Neither shows up as an error. They show up as a senior engineer's afternoon.
44
60
 
45
- So it guesses. It greps where a senior engineer would check the call graph. It re-derives on Tuesday what it worked out on Monday. And the knowledge that *would* brief it — who changed each file and why, what failed before, the conventions the team accreted is scattered across one tool for memory, another for the graph, a third for context, none of which can reach across the others.
61
+ The knowledge that would have stopped it — who calls this function, which pattern is load-bearing already exists. The whole market is built on getting that knowledge to the agent. And it falls into two shapes, both of which leak:
62
+
63
+ | What it does | The shape | Why it leaks |
64
+ |---|---|---|
65
+ | **Tells the agent things.** Memory stores, code-graph servers, context packers, rule files. | A tool the agent calls *when it remembers to.* | Optional context is optional. Agents skip the retrieval tool **~58% of the time even when explicitly told to use it** ([CodeCompass, 2026](https://arxiv.org/abs/2602.20048)). Advice it can ignore, it ignores. |
66
+ | **Checks the agent afterward.** Reviewers, linters, CI gates. | A pass over the diff *after the code is written.* | The break already happened. Now it's a comment on a pull request and a second round of work — not a change that never broke anything. |
46
67
 
47
- **unerr is the layer that ends the guessing.** One per-repo runtime, behind every MCP your agent already speaks, that carries your codebase's lived history and hands the agent that judgment the moment it starts so it lands on the right code without burning turns, and sees what a change will break before it breaks it.
68
+ There's a third shape, and almost no one ships it: **guidance wired into the moment of the edit, that the agent can't route around, and that re-anchors itself when the code moves so it never goes quietly stale.** Not a tool it chooses to consult. Not a review after the fact. A guardrail that fires *as it edits* — and stays true to the code because it's recomputed from the code, not from a doc that rots.
69
+
70
+ That's unerr. The agent doesn't have to ask. Before the edit lands, it already sees the callers it would break and the standard it's about to violate.
48
71
 
49
72
  | The old way | With unerr |
50
73
  |---|---|
51
- | The agent refactors a function blind to its 24 callers — 7 sites break silently. | **Cascade guard** reads the call graph *before* the edit; every caller is on screen first. |
52
- | Conventions and decisions live in `MEMORY.md` / `.cursorrules` you hand-maintain and re-paste each session. | **Anchored memory + drift detection** facts pin to the code and get a drift signal when it moves, instead of going silently stale. |
53
- | Five single-purpose MCP servers memory, graph, compressor that can't reach across each other. | **One runtime** so cascade guard, convention drift, and loop breaker fire on joins no point tool can make. |
74
+ | The agent changes a function without reading its 24 callers — 7 sites break silently. | **Cascade guard** puts the call graph in front of the edit *before it runs* every caller on screen, no asking required. |
75
+ | You wrote the rule in `.cursorrules`. The agent acknowledged it, then ignored it once context filled up. | **Anchored rules** surface the standard the instant the agent touches that scope and re-anchor when the code moves instead of going stale. |
76
+ | A rule or spec stays confident long after the code moved out from under it. Nothing recomputes it. | Every fact is pinned to a live entity in the graph. When the code moves, the fact **fails loud** instead of staying silently wrong. |
54
77
 
55
78
  ---
56
79
 
57
80
  ## The pains this fixes
58
81
 
59
- You've felt all four of these in the last 48 hours:
82
+ You know this feeling, and it gets *worse* as the repo grows, not better:
60
83
 
61
- - Claude is brilliant for 20 minutes, then hallucinates a duplicate component and forgets the styling rules you set five turns ago.
62
- - More time spent writing `MEMORY.md`, updating `.cursorrules`, and pasting session summaries than writing code.
63
- - The agent reads a 2,000-line file to find a 5-line function, then still doesn't know that function has 24 callers in six other files.
64
- - You don't trust the agent to refactor anything important. It treats your codebase like a flat string of text locally correct, globally wrong.
84
+ - **You're babysitting it.** You can't fire-and-forget, because the one time you look away is the time it quietly breaks something load-bearing. You've become its scheduler and its safety net at once.
85
+ - **You don't trust it to touch anything important.** It treats your codebase as a flat wall of text — locally correct, globally wrong — so the load-bearing changes still land on you.
86
+ - **The rule you wrote gets acknowledged, then dropped.** A few turns later the context fills up and your `.cursorrules` line may as well not exist.
87
+ - **Approval fatigue.** You approve so many reasonable edits that the dangerous one slides through the hundredth confirmation looks exactly like the first.
65
88
 
66
- These aren't four problems. They're one: **your agent acts on your codebase without ever having been briefed on it.** It greps where a senior engineer would check the call graph, and it relearns on Tuesday what it worked out on Monday.
89
+ These aren't four problems. They're one: **the agent acts on a codebase it can't hold in its head, and nothing it can't bypass is watching the change.** You babysit because there's no guardrail it can't skip. unerr is that guardrail so you can look away.
67
90
 
68
91
  ---
69
92
 
@@ -71,16 +94,16 @@ These aren't four problems. They're one: **your agent acts on your codebase with
71
94
 
72
95
  | You feel | What unerr does |
73
96
  |---|---|
74
- | **Trust returns.** The agent runs for an hour without you watching. | Every edit is preceded by a graph lookup. All 24 callers are visible *before* it touches the function. Refactors stop rippling silently. |
75
- | **The babysitter tax disappears.** You delete `MEMORY.md` and `.cursorrules`. | A local fact store remembers what you decided, what failed, and the conventions the team accretedwith decay-adjusted confidence. Open the laptop on Tuesday and the agent already knows what you decided on Monday. |
76
- | **The agent stays sharp at turn 50.** | `file_read({entity})` returns 200 lines instead of 3,000. Shell output is compressed 93% on average. The context window stays uncluttered, so the model isn't fighting "lost in the middle." |
77
- | **Tool sprawl dies.** | One graph, one set of tools, project-aware routing. Five MCP servers no longer compete for the agent's attention. |
97
+ | **You stop babysitting.** The agent runs for an hour and you're not bracing for a silent break. | Every edit is preceded — automatically — by a graph lookup. All 24 callers are visible *before* it touches the function. The guardrail fires whether or not the agent thought to ask. |
98
+ | **Your rules finally get honored.** The standard you set is applied at the edit, not acknowledged and forgotten. | unerr pins each rule and decision to the file or entity it governs and surfaces it the instant the agent touches that scope then re-anchors it when the code moves. Keep your `.cursorrules` and specs; unerr makes sure they're actually applied. |
99
+ | **It stops thrashing.** No more watching it retry the same broken fix three times. | A **loop breaker** watches the timeline and stops the agent re-trying a change that already failed twice before it burns your turn and your patience. |
100
+ | **The agent stays sharp at turn 50.** | `file_read({entity})` returns 200 lines instead of 3,000; shell output is trimmed automatically. The window stays uncluttered, so the model isn't fighting "lost in the middle." |
78
101
 
79
- **What it looks like in your chat** — before the Edit tool runs, unerr injects this into the agent's context:
102
+ **What it looks like in your chat** — before the Edit tool runs, unerr injects this into the agent's context, on its own:
80
103
 
81
104
  > ⚡ unerr · cascade guard: editing `src/payments/gateway.ts` changes a signature with callers that must be updated in the same change — `processPayment`: **24 callers at risk across 6 files** (19 source, 5 test). Call `get_references({key:'processPayment', direction:'callers'})` and update every caller before finishing.
82
105
 
83
- The outcome you get is **agents that behave like senior engineers** — checking dependencies before editing, remembering project history, refusing to thrash on a function they've already failed on three times.
106
+ The outcome: **agents that behave like senior engineers** — checking dependencies before editing, honoring the standard, and refusing to thrash on a function they've already failed on three times.
84
107
 
85
108
  ---
86
109
 
@@ -95,19 +118,19 @@ The outcome you get is **agents that behave like senior engineers** — checking
95
118
 
96
119
  Two places unerr shows up so you know it's working — inside the chat, and in a browser.
97
120
 
98
- **Inside the chat.** Every coding turn opens with one line naming what unerr loaded ("loaded a convention you wrote yesterday for `src/payments/gateway.ts`…") and closes with one line totalling what it saved you ("this turn: 2 catches · ≈ 4.2k tokens saved · +5 turns of headroom this session"). Catches are *named, countable events*, not a ratio.
121
+ **Inside the chat.** Every coding turn opens with one line naming what unerr loaded ("loaded a convention you wrote yesterday for `src/payments/gateway.ts`…") and closes with one line totalling what it caught and saved ("this turn: 2 catches · ≈ 4.2k tokens saved · +5 turns of headroom this session"). Catches are *named, countable events*, not a ratio.
99
122
 
100
- **In a browser.** A live dashboard at `http://localhost:9847` reads from the same store the agent reads from over MCP — the graph it navigates, the facts it remembers, the tokens it didn't have to chew through, and the score showing which of those facts actually shaped the next answer.
123
+ **In a browser.** A live dashboard at `http://localhost:9847` reads from the same store the agent reads from over MCP — the graph it navigates, the facts it remembers, the breaks it caught, and the score showing which of those facts actually shaped the next answer.
101
124
 
102
125
  <p align="center">
103
126
  <img src="https://unerr.dev/open-cli/screenshots/end-of-turn-receipt.png" alt="unerr end-of-turn receipt — tokens saved and headroom kept open this turn" width="380" />
104
127
  <img src="https://unerr.dev/open-cli/screenshots/end-of-turn-receipt-2.png" alt="unerr end-of-turn receipt — named, countable catches totalled at the close of a turn" width="380" />
105
- <br/><sub><strong>End-of-turn receipt</strong> · every coding turn closes with one line totalling what unerr saved you — named, countable catches, not a ratio.</sub>
128
+ <br/><sub><strong>End-of-turn receipt</strong> · every coding turn closes with one line totalling what unerr caught and saved you — named, countable catches, not a ratio.</sub>
106
129
  </p>
107
130
 
108
131
  <p align="center">
109
132
  <img src="https://unerr.dev/open-cli/screenshots/dashboard.png" alt="unerr dashboard — live overview" width="300" />
110
- <br/><sub><strong>Dashboard</strong> · live overview — active sessions, recent tool calls, tokens the agent skipped this turn.</sub>
133
+ <br/><sub><strong>Dashboard</strong> · live overview — active sessions, recent tool calls, breaks caught, tokens the agent skipped this turn.</sub>
111
134
  </p>
112
135
 
113
136
  <p align="center">
@@ -153,9 +176,9 @@ Install multiple agents in the same repo — each writes its own config. Idempot
153
176
 
154
177
  ### 3. Restart your IDE
155
178
 
156
- Close and reopen your IDE (or start a new chat session). Your agent picks up unerr through MCP — graph-backed tools, persistent memory, shell compression all available immediately.
179
+ Close and reopen your IDE (or start a new chat session). Your agent picks up unerr through MCP — graph-backed tools, persistent memory, the edit-time guardrail all available immediately.
157
180
 
158
- > **Dashboard:** <http://localhost:9847> — open any time to watch unerr's operational intelligence at work in real time.
181
+ > **Dashboard:** <http://localhost:9847> — open any time to watch unerr at work in real time.
159
182
 
160
183
  > Need manual setup or any other MCP client? `unerr install --show-instructions <agent>` prints copy-pasteable steps.
161
184
 
@@ -163,50 +186,54 @@ Close and reopen your IDE (or start a new chat session). Your agent picks up une
163
186
 
164
187
  ## Who it's for
165
188
 
166
- - **Vibe coders.** The thing that stops your app from breaking on turn 30 when the AI gets confused.
167
- - **Solo builders.** The continuous thread. Switch from Claude Code in the terminal to Cursor in the IDE your project memory comes with you.
168
- - **Senior / staff engineers.** The dependency graph, prior incidents, and team conventions a human engineer would already carry in their head fed to AI on every edit.
189
+ - **Engineers on large, existing codebases.** The dependency graph, the load-bearing patterns, and the prior incidents a senior engineer carries in their head — handed to the agent before every edit, so it stops breaking callers it never read.
190
+ - **Teams with conventions worth enforcing.** The standard you agreed on once, applied every time the agent touches that scope — no `.cursorrules` file to hand-maintain, re-paste, or merge-conflict over, and no hoping the agent remembers to look.
191
+ - **Solo builders shipping into a codebase that's already grown.** The continuous thread across tools switch from Claude Code in the terminal to Cursor in the IDE and the graph, rules, and history come with you, instead of relearning the repo every session.
169
192
 
170
193
  ---
171
194
 
172
- ## Why one runtime, not five separate tools
195
+ ## Why a guardrail has to be one runtime, not five tools
173
196
 
174
- **unerr is the layer your agents share — sitting *behind* every MCP they already speak.** Every coding agent on your machine Claude Code, Cursor, Windsurf, Antigravity speaks MCP. MCP carries tool calls; it does not carry context. Without unerr, every agent rebuilds your codebase's dependency graph, conventions, and prior decisions from scratch every session, by reading files blindly. With unerr, all of them read the same per-repo runtime over MCP, so your project's graph, memory, and guardrails carry across sessions *and* across IDEs.
197
+ A guardrail the agent *can't skip* can't be a tool the agent chooses to call. That's the whole reason unerr is one local runtime sitting *behind* the MCP every agent already speaksnot a fifth server in the agent's tool list.
175
198
 
176
- The adjacent space already has strong point tools. unerr's job is not to out-feature any of them in their laneit's to be the single per-repo runtime that joins them.
199
+ Every coding agent on your machine Claude Code, Cursor, Windsurf, Antigravity — speaks MCP. MCP carries tool calls; it does not carry context, and it does not fire anything on its own. So a memory server, a graph server, and a compressor sit there *waiting to be invoked* and an agent under context pressure skips them. unerr instead intercepts at the moment that matters the read, the edit — and injects the one scoped thing that's relevant, automatically. The agent can't forget to call something that isn't waiting to be called.
177
200
 
178
- | Layer | Where point tools live | What unerr adds |
179
- |---|---|---|
180
- | Memory across sessions | dedicated memory tools | Memory tied to the *current* state of the code — facts get drift signals when the file they're about moves. |
181
- | Code-graph navigation | dedicated code-graph tools | The graph is read *before every file read* — surgical context instead of 3,000-line dumps. |
182
- | Output compression | dedicated compression tools | Compression is fed through the same MCP runtime as the graph and memory, not a separate tool the agent has to remember to invoke. |
183
- | Convention enforcement | `.cursorrules`, CLAUDE.md hand-maintained | Conventions auto-detected from ≥70% adherence in the code. No file to maintain. |
201
+ That only works if the pieces live in **one** process. The guardrails worth having each fire on a *join* no single tool can make:
184
202
 
185
- We deliberately don't ship a feature-by-feature checkmark matrix against the depth leaders on each lane — that's the trap. A dedicated memory tool will out-memory us on memory depth; a dedicated code-graph tool will out-graph us on graph aesthetics; a dedicated compressor will out-compress us on shell compression simplicity. The runtime is the join across all four lanes — not the depth on any one.
203
+ - **Cascade guard** needs the call graph *and* the edit-intent ledger on the same process, at the same instant.
204
+ - **Drift** needs memory that's anchored to a live graph — so the fact knows the moment its code moved.
205
+ - **Convention drift** needs the auto-detected pattern store *and* the new-code stream in the same memory space.
206
+ - **Loop breaker** needs the full timeline of what the agent already tried.
186
207
 
187
- Three numbers behind the runtime:
188
-
189
- - **~84%** of an AI coding agent's tokens are tool output, mostly file reads ([JetBrains, NeurIPS 2025](https://blog.jetbrains.com/research/2025/12/efficient-context-management/)) — unerr intercepts at the read layer, so attention isn't diluted.
190
- - **Tool-selection accuracy collapses 58% → 26% as MCP tools go from 9 to 51** ([LangChain ReAct benchmark](https://blog.langchain.com/react-agent-benchmarking/)) — unerr is one MCP runtime instead of five, freeing the agent's tool-selection budget. Anthropic itself acknowledged this in Jan 2026 by shipping [MCP Tool Search](https://www.anthropic.com/engineering/code-execution-with-mcp) to hide tool definitions until queried.
191
- - **0** LLM calls per query in the core — facts, conventions, drift signals, and graph lookups are all algorithmic. No API keys, no per-turn inference cost, no telemetry.
192
- - **86–90%** of an agent's code-navigation tokens removed in head-to-head benchmarks vs grep+read — real tokenizer, fidelity-gated, reproducible on any repo ([benchmarks](./benchmarks/README.md)).
208
+ These aren't features you can buy individually and bolt together. They're emergent properties of one runtime — and they're exactly what turns "context the agent might read" into "a guardrail it can't skip."
193
209
 
194
210
  ---
195
211
 
196
212
  ## How the runtime works
197
213
 
198
- One local process per repo. Four slices, joined deterministically — *the joins are the product, not the slices.* Point tools own one slice each. None of them can ship the joins without becoming a per-repo runtime themselves.
214
+ One local process per repo. Four mechanisms, joined deterministically — the **mechanisms** are how; the **guardrail** is what you get.
199
215
 
200
- | Slice | What's inside | What the join enables |
216
+ | Mechanism (the how) | What's inside | What it powers (the what) |
201
217
  |---|---|---|
202
- | **Live code graph** | CozoDB · tree-sitter ASTs · SCIP-verified call graphs · 18+ languages · <5ms queries | Read *before every file read*. The agent opens 50 targeted lines and a caller list — not 3,000 lines and a guess. |
218
+ | **Live code graph** | CozoDB · tree-sitter ASTs · SCIP-verified call graphs · 18+ languages · <5ms queries | The agent opens 50 targeted lines and a caller list — not 3,000 lines and a guess. Read *before every file read*, so cascade guard knows what an edit breaks. |
203
219
  | **Anchored memory** | Typed facts · conventions auto-detected at ≥70% adherence · decay-adjusted confidence | Every fact is pinned to a file or entity in the graph. When the code moves, the fact gets a **drift signal** — never silent staleness. |
204
- | **Context delivery** | Shell output compression (93% overall, 645+ command classifiers) · Web fetches (5–10× via Defuddle + BM25) · Entity-targeted file reads | Compression, graph, and memory share one process — the agent doesn't have to remember which tool to invoke for which kind of content. |
205
- | **Behaviour modules** | cascade guard · convention drift · loop breaker · session continuity · auto-doc · change narrative · architecture guard | Each guardrail fires on a *join* cascade-guard reads the graph before the edit, convention-drift compares new code against memory, loop-breaker watches the timeline. None of these are reachable from a single point tool. |
220
+ | **Context delivery** | Shell output compression (645+ command classifiers) · web fetches (5–10× via Defuddle + BM25) · entity-targeted file reads | The relevant slice arrives automatically at the read — the agent never has to remember which tool to invoke for which content. |
221
+ | **Behaviour modules** | cascade guard · convention drift · loop breaker · session continuity · auto-doc · change narrative · architecture guard | Each guardrail fires on a join of the three above, *at the moment of the edit* not as a tool the agent chose, not as a review after the fact. |
222
+
223
+ **The unifying point.** Drift detection requires memory anchored to a live graph. Cascade guard requires the graph and the edit-intent ledger on one process. Convention drift requires the pattern store and the new-code stream in the same memory space. Spread these across five disconnected MCP servers and none of them can fire — they can only sit and wait to be called, which is the failure mode this whole thing exists to fix. That's the difference between a stack of tools and a guardrail.
206
224
 
207
- **The unifying point.** Drift detection requires memory anchored to a live graph. Cascade-guard requires the graph and the edit-intent ledger on the same process. Convention-drift requires the auto-detected pattern store and the new-code stream in the same memory space. These aren't "features" you can buy individually — they're *emergent properties of the runtime*, only available when all four slices live in one per-repo process.
225
+ ---
226
+
227
+ ## Fewer tokens, as a side effect
228
+
229
+ unerr was built to stop bad changes, not to save tokens. But a guardrail that only ever hands over *the one scoped fact that matters* — the rule for the entity in front of the agent, 50 lines instead of 3,000 — spends far fewer tokens almost by accident. So you get this for free:
230
+
231
+ - **86–90%** of an agent's code-navigation tokens removed in head-to-head benchmarks vs grep+read — real tokenizer, fidelity-gated (any "saving" that lost the answer is discarded), reproducible on any repo. [See the benchmarks →](./benchmarks/README.md)
232
+ - **~84%** of an AI coding agent's tokens are tool output, mostly file reads ([JetBrains, NeurIPS 2025](https://blog.jetbrains.com/research/2025/12/efficient-context-management/)) — unerr intercepts at the read layer, so the window isn't diluted.
233
+ - **Tool-selection accuracy collapses 58% → 26% as MCP tools go from 9 to 51** ([LangChain ReAct benchmark](https://blog.langchain.com/react-agent-benchmarking/)) — unerr is one runtime instead of five servers, so it doesn't eat the agent's tool-selection budget. Anthropic itself acknowledged this in Jan 2026 by shipping [MCP Tool Search](https://www.anthropic.com/engineering/code-execution-with-mcp).
234
+ - **0** LLM calls per query in the core — facts, conventions, drift signals, and graph lookups are all algorithmic. No API keys, no per-turn inference cost, no telemetry.
208
235
 
209
- Five disconnected MCP servers one for memory, one for graph, one for compression, one for tracing, one for skills burn ~55K tokens of schemas just to *announce themselves* (Anthropic's own engineering example). They can't reach across each other to fire any of these guardrails. That's the difference between a stack and a runtime.
236
+ The point was never the token number. It's that the agent lands on the right code, sees the right guardrail, and you stop paying in tokens *and* in afternoons for the changes it would otherwise have to undo.
210
237
 
211
238
  ---
212
239
 
@@ -240,7 +267,7 @@ One local DB per repo. Zero network calls. No API keys. No cloud. Your code neve
240
267
 
241
268
  Full module map and source-tree breakdown: **[ARCHITECTURE.md](./docs/ARCHITECTURE.md)**.
242
269
 
243
- **Design principles** — zero network calls; stdout is sacred (MCP JSON-RPC only, everything else to stderr); <5 ms query responses; first useful output <5 s (shallow index first, deep enrichment in background); graceful degradation (the agent still works if unerr is down, you just lose the operational intelligence layer).
270
+ **Design principles** — zero network calls; stdout is sacred (MCP JSON-RPC only, everything else to stderr); <5 ms query responses; first useful output <5 s (shallow index first, deep enrichment in background); graceful degradation (the agent still works if unerr is down, you just lose the guardrail layer).
244
271
 
245
272
  **Tech stack** TypeScript (ESM) · CozoDB (Rust/NAPI) · web-tree-sitter (WASM) · MCP SDK · Ink (React CLI) · React + Vite (dashboard) · tsup · Vitest
246
273