Repository navigation
Releases
Releases · raulkivi/agent-cost-analyzer
Agent Cost Analyzer 1.2.0
Added
- The MCP server is listed in the official MCP Registry as
io.github.raulkivi/agent-cost-analyzer. - A Claude Code plugin in this repository registers the MCP server:
/plugin marketplace add raulkivi/agent-cost-analyzer, then/plugin install agent-cost-analyzer@agent-cost-analyzer. - The standalone app's header has a Connect an agent button: a dialog that shows how to register the MCP server with each client, with copy buttons. It changes nothing on its own.
- Each app release attaches a Claude Desktop bundle (
agent-cost-analyzer-<version>.mcpb) that installs the MCP server when opened.
AI Credits Analyzer 1.0.0
Added
- Prices for Claude Fable 5.2, Gemini 4 Carbon and Gemini 4 Hydrogen, from the Copilot price table that GitHub publishes (checked 2026-11-16). Gemini 4 Carbon is a promotional price until 2026-12-01.
- A GitHub Copilot Chat session's turn now lists the tool calls of the subagent runs it started, after the main agent's calls, each marked with the run it belongs to, so the tool-call counts and failures include them. A run whose log is missing, unreadable or empty, a run started by another subagent, and a subagent tool that writes no log are named with the reason their tool calls are unknown, never counted as zero. A call that appears in both the main log and a run's log is counted once.
- A "Tools loaded but never used" row in the Savings tab and in the chat view now lists its tools: each with its definition tokens per request (an estimate, "about N") and "loaded in N sessions". It shows the five largest and "and N more (about X)", where X is the row's total minus the five.
get_savings_reportcarries up to 20 tools and prints the first five. - Each session in a savings row now shows its date and its own saving over the window, with the row's label ("up to", a range or Simulated), in the Savings tab and in the chat view; a session that could not be priced says "Not priced" with the reason once, and one without a start says "date unknown".
get_savings_reportcarries the same two fields per shown session (no title) and prints them as<id> (<date>, <saving>). The tool's Simulated rows no longer have arankand are bullets, not numbers, in its text, and the chat view reads every coverage gap as "N not analysed: reason". - Each session in a savings row shows its title first, then its id and date under it, in the Savings tab (the standalone app and the extension). The title is looked up in the browser from the session list the Dashboard loads, shown as text on one line (cut at 120 characters, in full on hover); a session without a title shows the id alone. Titles are not in the report,
get_savings_report, the chat view or Copy. - A session whose model has no known price channel now gets its "Tools loaded but never used" row with "Not priced" and the reason, instead of being left under Not analysed.
- The MCP tool-call audit tools tell the main agent's calls from subagents'.
get_session_auditandget_audit_rollupreturnbyOrigin(andget_session_auditthesubagentRunswith the reason for any run that could not be read;get_audit_rollupsessionsWithUnreadSubagentRuns),list_tool_callsandget_turnmark each call'sorigin, andlist_findingsnames the run behind a finding.list_tool_callsandlist_findingstake anoriginfilter, andlist_tool_callsarunIdfilter; with neither, both are returned and the totals are the true totals. A subagent call's own arguments (its prompt) are not shown. - The MCP App's tool-call audit view shows subagent runs. A session with subagent runs gets "incl. N calls from M subagent runs" under the Tool calls figure, a Calls by agent split, an All / Main agent / Subagents filter for the figures, and a Subagent runs list with each run's calls, failures and unknown outcomes. A run whose calls could not be read says why and shows Unknown, never 0; a run with a flag keeps its figures. The date-range audit adds the same sub-line and the number of sessions with unread runs, and user-canceled calls are listed apart from failures. The findings view names the subagent run behind a finding ("by subagent run Agent-Explore (turn 4)").
- The Tool audit tab and the All sessions audit show subagent tool calls. The Tool audit tab's Tool calls tile says "incl. N calls from M subagent runs" ("from K of M" when some runs could not be read), with a Calls by agent split, an Agent filter (All, Main agent, Subagents), an Agent column in the calls table, a Subagent runs table (an unread run shows its reason, never 0) and a Canceled count for calls the user stopped (not failures). The All sessions audit splits its totals into main agent and subagent runs, adds a Subagent calls column ("n/a" where the log has no subagents) and counts the sessions with unread runs. The turns table's Calls cell says how many calls came from subagent runs, the Explanation panel's subagent block gives each run its calls and failures, and a finding from a subagent call names its run.
- A run the analyzer placed by time instead of by a link to its start call (an orphan subagent log, an
explore_subagentrun matched by its time window, a Claude Code run matched by its first timestamp) now says "matched by time" in the tool-call audit: the Subagent runs table, the MCP App's runs list, and thematchedByTimefield ofget_session_audit'ssubagentRuns.
Changed
- A tools row's "and N more" shows "(about X)", the rest's tokens, only when every listed tool was loaded in every session of the row; otherwise it shows just "and N more". A tool's figure is its average over the sessions that loaded it, so the difference would misstate the rest.
- A cache-break row (idle expiry, no idle gap, model switch) now leaves a session with any break that has no price out of its amount and lists it under the row's coverage with the reason, as the failed-calls row does; the session's own saving reads "Not priced". Before, the amount counted such a session's priced breaks only.
Fixed
- In the chat, a Claude Code session's cost view now gains a Priced cost tile beside AI Credits, as in the web app; AI Credits reads "n/a" ("Claude Code has no AI Credits") instead of "Unknown". Its Top turns table reads "n/a" in the AI Credits column. An unknown priced cost still reads "Unknown" with its reason, never $0. Copilot sessions look as before.
- The Tool audit tab's status chips (Succeeded, Failed and the rest) count only the calls the Agent filter shows. With Subagents selected, the chips no longer include the main agent's calls, so each chip matches the rows it toggles.
- A savings row's coverage line now lists each reason once, with the summed count, instead of the same reason twice with two counts (for example "13 not analysed: The log records no tool inventory for this session." where it read "5 not analysed: ... · 8 not analysed: ..."). It showed up when sessions from more than one log source shared the reason, in the Savings tab, the chat view and
get_savings_report. - Reading a Copilot or VS Code session store no longer fails with "database is locked" while VS Code or Copilot is writing to it. The read now waits up to 5 seconds for the other app's commit instead of failing at once; the analyzer still never writes those files.
- A session in a savings row whose VS Code Agent Host title only repeated its first prompt no longer shows "Session " as its title: the row shows the id alone, as for any untitled session.
- VS Code Agent Host sessions that some VS Code Insiders builds cataloged under an
ahp-session:/URI, instead ofcopilotcli:/, were missing everywhere: the sessions list, costs and savings. They now appear like any other Agent Host session, with their turns, AI Credits and turn text. A session cataloged under both URIs shows once. - Savings action text now agrees with its counts: "1 session", "1 tool", "1 day", "1 break", "1 time", "1 token" and "1 minute" instead of "1 sessions" and so on.
- The extension now ships third-party license notices (
THIRD-PARTY-NOTICES.mdin the .vsix) for the npm packages it bundles, with their license texts. get_savings_report's text answer says "1 of 1 session" and "1 session." instead of "1 of 1 sessions" and "1 sessions." in a row's coverage and in the lists of sessions not checked or left out.get_savings_report's description now tells the model that each listed session carries its start and its own saving over the window (not per week), and that an unused-tools row lists up to 20 of its tools with estimated tokens.
Agent Cost Analyzer 1.1.0
Added
agent-cost-analyzer mcp install,uninstall,statusandconfigregister the MCP server with Claude Code, Claude Desktop, Cursor, VS Code, VS Code Insiders, Codex and Gemini CLI.installanduninstallshow each step first and ask before applying (--dry-runstops after the plan,--yesskips the question);statusreads without writing;configprints the command, JSON or TOML to paste yourself.- The app prints how to register its MCP server (
npx agent-cost-analyzer mcp install) when it starts in a terminal.
Agent Cost Analyzer 1.0.0
Added
- Prices for Claude Fable 5.2, Gemini 4 Carbon and Gemini 4 Hydrogen, from the Copilot price table that GitHub publishes (checked 2026-11-16). Gemini 4 Carbon is a promotional price until 2026-12-01.
- A GitHub Copilot Chat session's turn now lists the tool calls of the subagent runs it started, after the main agent's calls, each marked with the run it belongs to, so the tool-call counts and failures include them. A run whose log is missing, unreadable or empty, a run started by another subagent, and a subagent tool that writes no log are named with the reason their tool calls are unknown, never counted as zero. A call that appears in both the main log and a run's log is counted once.
- The standalone app's header, in Analyze mode, shows a monthly AI Credits allowance gauge: the month's total of the allowance, with a meter and the level in words ("near limit", "over"; "at least" when some sessions have no cost data; "unknown" never reads as 0). Click or Enter opens a details dialog (period, coverage, split by harness, what is not counted); Change allowance opens a settings dialog with Remove allowance. The extension does not show it.
- The savings report now lists unused tools for pi sessions: each tool a pi session declared and never ran, with its definition tokens per request (an estimate) and the requests it was declared for, including a tool pi dropped or added mid-session. pi on a local model has no price, so its rows say "Not priced" with the reason and rank last; their tokens and requests are shown, never a made-up dollar figure. A branch of a forked pi file counts the requests it shares with an earlier branch once. Sessions without a system message in the file (only the logger's record) stay under Not analysed.
- MCP tool
get_monthly_allowancein the standalone server: the AI Credits used so far this billing period against the monthly allowance set in the app, the same answer as the app's header. Read-only, no arguments; a lower bound says "at least", and unknown is never 0. The extension's MCP server does not have it. - Any session whose model has no known price channel, not only pi, now gets its unused-tools row with "Not priced" and the reason, instead of being left under Not analysed.
- The MCP tool-call audit tools tell the main agent's calls from subagents'.
get_session_auditandget_audit_rollupreturnbyOrigin(andget_session_auditthesubagentRunswith the reason for any run that could not be read;get_audit_rollupsessionsWithUnreadSubagentRuns),list_tool_callsandget_turnmark each call'sorigin, andlist_findingsnames the run behind a finding.list_tool_callsandlist_findingstake anoriginfilter, andlist_tool_callsarunIdfilter; with neither, both are returned and the totals are the true totals. A subagent call's own arguments (its prompt) are not shown. - The MCP App's tool-call audit view shows subagent runs. A session with subagent runs gets "incl. N calls from M subagent runs" under the Tool calls figure, a Calls by agent split, an All / Main agent / Subagents filter for the figures, and a Subagent runs list with each run's calls, failures and unknown outcomes. A run whose calls could not be read says why and shows Unknown, never 0; a run with a flag keeps its figures. The date-range audit adds the same sub-line and the number of sessions with unread runs, and user-canceled calls are listed apart from failures. The findings view names the subagent run behind a finding ("by subagent run Agent-Explore (turn 4)").
- Claude Code sessions now report their subagent runs beside the totals.
get_session_costand the Sessions list show the number of runs with their tokens and priced cost, since Claude Code records no AI Credits; an unknown figure keeps its reason, never 0. A session replay (replay_session,compare_models_for_session) covers the main agent only and its coverage lists the runs it left out with their tokens and cost; the measured cost keeps them. The savings report reads the main agent only and each row says "N subagent runs not analyzed". - A Claude Code session's turn now lists the tool calls of the subagent runs it started, after the main agent's calls, each marked with the run it belongs to, so the tool audit and the security findings include them. A run whose transcript is missing, not a regular file, unreadable, or nested beyond the read depth is named with the reason its tool calls are unknown, never counted as zero. A call a fork copied from its parent is counted once.
- The Tool audit tab and the All sessions audit show subagent tool calls. The Tool audit tab's Tool calls tile says "incl. N calls from M subagent runs" ("from K of M" when some runs could not be read), with a Calls by agent split, an Agent filter (All, Main agent, Subagents), an Agent column in the calls table, a Subagent runs table (an unread run shows its reason, never 0) and a Canceled count for calls the user stopped (not failures). The All sessions audit splits its totals into main agent and subagent runs, adds a Subagent calls column ("n/a" where the log has no subagents) and counts the sessions with unread runs. The turns table's Calls cell says how many calls came from subagent runs, the Explanation panel's subagent block gives each run its calls and failures, and a finding from a subagent call names its run.
- A run the analyzer placed by time instead of by a link to its start call (an orphan subagent log, an
explore_subagentrun matched by its time window, a Claude Code run matched by its first timestamp) now says "matched by time" in the tool-call audit: the Subagent runs table, the MCP App's runs list, and thematchedByTimefield ofget_session_audit'ssubagentRuns. - A Claude Code session's subagent runs now show in the web app and the MCP App's session cost view as tokens and a priced cost, never an AI Credits figure. The turns table marks the turn with a fork glyph (tooltip "incl. 2.04M tokens, $0.38 priced, from 2 subagent runs"), the Explanation panel lists each run's model, requests, tokens and priced cost beside the main agent's and the turn's, and the narrow layout's KPI tiles read "AI Credits n/a" with a Priced cost tile that carries the runs' sub-line. A run that could not be read shows dashes and its reason, and no total adds only what is known.
Changed
- A tools row's "and N more" shows "(about X)", the rest's tokens, only when every listed tool was loaded in every session of the row; otherwise it shows just "and N more". A tool's figure is its average over the sessions that loaded it, so the difference would misstate the rest.
- A cache-break row (idle expiry, no idle gap, model switch) now leaves a session with any break that has no price out of its amount and lists it under the row's coverage with the reason, as the failed-calls row does; the session's own saving reads "Not priced". Before, the amount counted such a session's priced breaks only.
- The MCP tool descriptions and server instructions now explain Claude Code subagent runs: a turn's tokens include them,
get_session_cost'ssubagentsgives tokens and priced cost, replay leaves them out, and an unread or unplaced run is unknown, never 0
Fixed
- A savings row's coverage line now lists each reason once, with the summed count, instead of the same reason twice with two counts (for example "13 not analysed: The log records no tool inventory for this session." where it read "5 not analysed: ... · 8 not analysed: ..."). It showed up when sessions from more than one log source shared the reason, in the Savings tab, the chat view and
get_savings_report. - In the chat, a Claude Code session's cost view now gains a Priced cost tile beside AI Credits, as in the web app; AI Credits reads "n/a" ("Claude Code has no AI Credits") instead of "Unknown". Its Top turns table reads "n/a" in the AI Credits column. An unknown priced cost still reads "Unknown" with its reason, never $0. Copilot sessions look as before.
- The Tool audit tab's status chips (Succeeded, Failed and the rest) count only the calls the Agent filter shows. With Subagents selected, the chips no longer include the main agent's calls, so each chip matches the rows it toggles.
- The replay summary of a Claude Code session with subagent runs now gives the runs' priced cost and tokens ("Their priced cost is $9.20 for 31.4M tokens.") instead of "Their cost is unknown: AI Credits are GitHub Copilot's own billing unit". A Copilot session keeps its AI Credits sentence. If the priced cost itself is unknown, the sentence says so with the reason, never 0.
- Reading a Copilot or VS Code session store no longer fails with "database is locked" while VS Code or Copilot is writing to it. The read now waits up to 5 seconds for the other app's commit instead of failing at once; the analyzer still never writes those files.
- A session in a savings row whose VS Code Agent Host title only repeated its first prompt no longer shows "Session " as its title: the row shows the id alone, as for any untitled session.
- A Claude Code session with a subagent transcript that no turn could be matched to no longer fails to load in clients that validate the answer: the turn that carries that session-level unknown keeps the main agent's own figures beside it, and the shared turn-usage schema now accepts that shape.
- Claude Code subagent runs are now counted in a session's tokens and cost. Each run is added to the turn whose
Agentcall started it, priced on its own model, and the turn keeps the main agent's own figures beside it. The Sessions list and the session detail read the same runs, so they agree. A run that is missing, unreadable or cannot be placed makes its turn (or, for a folder that cannot be read, the session) unknown with the reason, never lower. Runs of a branch you rewound are not counted, and a run...
AI Credits Analyzer 0.14.0
Added
- MCP tool
get_savings_report: the Savings tab's ranked opportunities for a model (kind, amount as a range or "up to" with its basis and price versions, coverage, session ids, fixed action text; names and reasons redacted and capped; at most 20 rows, Simulated cheaper-model rows only withincludeSimulated), andreview_costsstarts from it. - The savings report in the chat:
get_savings_reportshows its ranked changes as a view (a range or "up to" per week, coverage, the Simulated section, Not analysed). A row opens in place with its action text and sessions; a session opens its cost or its replay in place, with Back. It sends nothing to the chat.
Fixed
- A saving smaller than 0.1 AI Credits (or $0.01) a week now reads "under 0.1 AI Credits per week" ("under $0.01 per week") instead of 0.0–0.0, which looked like no saving. The same goes for the amount over the window and per session; a range that starts at 0.0 but reaches 0.1 or more is unchanged.
- Copilot Chat sessions whose stored agent name was a helper's (
panel/editAgent,summarizeConversationHistory) are now listed and counted. They are ordinary Chat-panel chats (the store takes the name from one request) and were left out of every view; sessions are now selected by surface (Chat panel), still leaving out the Copilot CLI, the Coding Agent and subagents. - An Agent Host session that VS Code titled with the opening of its first prompt (some are, in the catalog and in the session's
customTitle) now shows as untitled everywhere, so session lists, statements, exports and MCP answers such asget_statementno longer carry prompt text. A title you set yourself is kept. - Replaying a local Copilot Chat session on another model now follows the main agent's requests only. The todo-list agent's and the conversation summary's requests have prompts of their own; they used to be replayed as part of the main agent's cache, which made the self-check unavailable (a helper on another vendor's model) or fail next to them. They are now left out and listed in the replay's coverage with their AI Credits (
coverage.helperRequestsinreplay_sessionandcompare_models_for_session); their cost stays in the measured figures. - Learn's Vendors tab fits the panel's side column: one block per model instead of a six-column table, so names no longer wrap and the Misses column is no longer cut off. Sorting moves to a Sort by selector.
Changed
- The chat views' Analyse (session picker) and Ask about this session buttons now read Added to chat after a click, with the hint "Press send in the chat to run it.", since the hosts put the message in the chat input rather than sending it.
Agent Cost Analyzer 0.16.1
Added
- The npm package
agent-cost-analyser(British spelling) is now a working alias ofagent-cost-analyzer:npx agent-cost-analyserruns the same app, with the same options and themcpcommand. It has one dependency, the real package, and no install scripts, and is published through the same release workflow with provenance.
Agent Cost Analyzer 0.16.0
Added
- A "Tools loaded but never used" row now lists its tools: each with its definition tokens per request (an estimate, "about N"), averaged over the sessions that loaded it, and "loaded in N sessions". The Savings tab and the chat view show the five largest and "and N more (about X)", where X is the row's total minus the five.
get_savings_reportcarries up to 20 tools (tools,toolsTotal) and prints the first five. Under an exactmcp__<server>__group a tool shows its tool part; under an inferredmcp_group, its whole name. - Each session in a savings row now shows its date and its own saving over the window, with the row's label ("up to", a range or Simulated), in the Savings tab and in the chat view; a session that could not be priced says "Not priced" with the reason once, and one without a start says "date unknown".
get_savings_reportcarries the same two fields per shown session (no title) and prints them as<id> (<date>, <saving>). The tool's Simulated rows no longer have arankand are bullets, not numbers, in its text, and the chat view reads every coverage gap as "N not analysed: reason". - Each session in a savings row shows its title first, then its id and date under it, in the Savings tab (the standalone app and the extension). The title is looked up in the browser from the session list the Dashboard loads, shown as text on one line (cut at 120 characters, in full on hover); a session without a title shows the id alone. Titles are not in the report,
get_savings_report, the chat view or Copy. - The npm package
agent-cost-analyzer, from app 0.16.0:npx agent-cost-analyzerstarts the app on http://127.0.0.1:5199 and opens your browser, with no clone. A running instance is reused, and a busy port moves to the next free one;--port <n>takes one port only,--no-openskips the browser,--helpand--versiondo what they say. It needs Node.js 22.18 or newer and has no dependencies and no install scripts. It is published from the release workflow through npm trusted publishing, with a provenance attestation. agent-cost-analyzer mcpruns the read-only MCP server over stdio from the same package, with the interactive views included. Register it withnpx -y --prefer-offline agent-cost-analyzer mcp, or install it globally for offline use.
Fixed
- Savings action text now agrees with its counts: "1 session", "1 tool", "1 day", "1 break", "1 time", "1 token" and "1 minute" instead of "1 sessions" and so on, in every change's wording.
Agent Cost Analyzer 0.15.0
Added
- The chat views label figures priced from a price version you added "your prices", as the app does: cost over time and project cost, compare, cost change, the statement, the estimate and the savings rows. Billed and reported figures never carry it. The VS Code extension reads only the bundled price book, so its views never show the label.
- Figures priced from a price version you added (
channels.priceVersionsinprojects.json) are labelled "your prices" in the app, with the version ids in the label's name and tooltip: the Dashboard's Cost tile, Cost over time, Compare, the Project tab, statement lines and the estimate, and the Savings rows. Billed and reported figures never carry it. - MCP tool
get_savings_report: the Savings tab's ranked opportunities for a model (kind, amount as a range or "up to" with its basis and price versions, coverage, session ids, fixed action text; names and reasons redacted and capped; at most 20 rows, Simulated cheaper-model rows only withincludeSimulated), andreview_costsstarts from it. - The savings report in the chat:
get_savings_reportshows its ranked changes as a view (a range or "up to" per week, coverage, the Simulated section, Not analysed). A row opens in place with its action text and sessions; a session opens its cost or its replay in place, with Back. It sends nothing to the chat.
Fixed
- A saving smaller than 0.1 AI Credits (or $0.01) a week now reads "under 0.1 AI Credits per week" ("under $0.01 per week") instead of 0.0–0.0, which looked like no saving. The same goes for the amount over the window and per session; a range that starts at 0.0 but reaches 0.1 or more is unchanged.
- Copilot Chat sessions whose stored agent name was a helper's (
panel/editAgent,summarizeConversationHistory) are now listed and counted. They are ordinary Chat-panel chats (the store takes the name from one request) and were left out of every view; sessions are now selected by surface (Chat panel), still leaving out the Copilot CLI, the Coding Agent and subagents. - An Agent Host session that VS Code titled with the opening of its first prompt (some are, in the catalog and in the session's
customTitle) now shows as untitled everywhere, so session lists, statements, exports and MCP answers such asget_statementno longer carry prompt text. A title you set yourself is kept. - The standalone app reads a VS Code install only when its folder has a
Userfolder inside. A leftover~/.config/Codeholding only an "Azure Accounts" folder is no longer picked over a real install, and no longer shows an empty session list. - Replaying a local Copilot Chat session on another model now follows the main agent's requests only. The todo-list agent's and the conversation summary's requests have prompts of their own; they used to be replayed as part of the main agent's cache, which made the self-check unavailable (a helper on another vendor's model) or fail next to them. They are now left out and listed in the replay's coverage with their AI Credits (
coverage.helperRequestsinreplay_sessionandcompare_models_for_session); their cost stays in the measured figures. - The standalone app's Overview passes the axe accessibility check: the AI Credits by day columns' names match their visible labels, and the session links in Most expensive sessions are 24px targets.
- A billing statement now covers every turn of its period: it read only the 50 latest sessions per source that started within a day of the period, so a busy month or a session started before the month left turns off the statement without saying so. It now reads up to 500 sessions per source, and when a source has more it says how many sessions were not read (in the Billing tab,
get_statementand the CSV and Markdown exports), never a silently smaller statement. - pi sessions' system prompt now matches what pi sent when the
pi-system-prompt-loggerextension is installed (exact text and size; pi's own sections as the breakdown, the extensions' additions as their own part), and says so when it is incomplete. - pi sessions: the turn inspector marks a response or tool result that pi later left out of the model's context ("omitted from later context") or replaced ("replaced in later context"). pi 1.5.2 does this itself when it recovers from a truncated or overflowing response (
context_edit). Its tokens still count, as pi billed them. The inspector also no longer shows these entries, or a compaction between two rounds, as raw JSON: a compaction shows its summary text. - Learn's Vendors tab fits its column: one block per model (name, per-turn strip, labelled figures) instead of a six-column table, so model names stay on one line and nothing scrolls sideways. Sorting moves to a Sort by selector.
Changed
- The chat views' Analyse (session picker) and Ask about this session buttons now read Added to chat after a click, with the hint "Press send in the chat to run it.", since the hosts put the message in the chat input rather than sending it.
AI Credits Analyzer 0.12.0
Added
- Savings view in the panel. Savings, after Dashboard in the panel's view switch, ranks changes
that would cost less over your GitHub Copilot sessions: tools loaded but never used, cache lost after an
idle gap or with no idle gap, a model switched mid-session, repeated failed tool calls, and with
Include simulated a cheaper model. Each row shows a range or an "up to" figure per week in AI
Credits, its coverage, and fixed action text with Copy text (one line, at most 600 characters). There
is no total, and what could not be checked is listed under Not analysed. Claude Code and pi savings
are in the standalone app. The report changes no setting or file. - Monthly AI Credits allowance gauge. Set
aiCreditsAnalyzer.monthlyAllowance.aiCreditsto
your plan's monthly allowance, or run Set Monthly AI Credits Allowance to open it in the
Settings editor (the command changes no setting). The status bar then shows the month to date
against the allowance, with a warning near the limit (80 % by default,warnAtPercent) and when
over. The Sessions view's first row shows the same figures with the days left. Unknown sessions
are never counted as 0, so the figure reads "at least" when some sessions have no cost data. The
gauge shows no notification and changes no setting. - Subagent line in the Sessions tooltip. Hover a classic session to see "incl. X AI Credits from N
subagent runs". - The MCP App turn detail lists subagent runs. Opening a turn in the session cost view shows each
run's agent, model, requests and AI Credits, or the reason they are unknown.
Changed
- Totals now include classic Copilot Chat subagent runs. A classic chat that called a subagent
(runSubagent,searchSubagent) wrote the run to its own debug log, which the analyzer left out.
Each run's AI Credits are now added to the turn that started it, in the turns table, the session
total, the dashboard and the MCP tools. Some totals rise. Some become unknown, with a reason: a
run's log is missing or can't be read, a run can't be matched to a turn, or a session has more than
200 subagent logs. A run whose log no event names is matched by time and labelled so. - Per-model figures, statements and price-change findings credit each run's model. A turn that
ran subagents on another model now counts under each model it used. Statements issued earlier may
show a diff when re-run, because the re-run includes the runs. - Replay compares the main agent. A replayed turn is compared on the main agent's own usage; the
coverage line lists the subagent runs it left out. - Denser panel: smaller spacing, 28px table rows, one toolbar row on the Dashboard, compact KPI tiles with the cost explanation in a footnote, a trend chart that fills its panel and one-line empty charts.
- Bar labels and code-style text now use the editor font instead of the browser's default monospace.
Fixed
- Opened-turn detail text contrast in VS Code Dark Modern. The opened-turn detail in the MCP App
view now sits on a neutral row background, so its text is readable. - Copilot logs whose records hold the characters U+2028 or U+2029, or a lone carriage return, inside a message are no longer split and dropped, so those sessions show all their requests and costs.
- In a narrow panel the All sessions audit's tables scroll inside their panels instead of making the view wider.
Agent Cost Analyzer 0.14.0
Added
- Savings tab. A ranked list of changes that would cost less: tools loaded but never used, cache lost
after an idle gap or with no idle gap, a model switched mid-session, repeated failed tool calls. Each row
shows a range or an "up to" figure per week, its coverage and the fixed action text with Copy (one line,
at most 600 characters); one list per currency (AI Credits, US dollars), never converted, and no total,
because the opportunities overlap. Rows carry "estimate", "inferred" or "Not priced" where they apply.
A Simulated section holds replays on a cheaper model, apart and unranked and off until "Include
simulated" is ticked, and "Not analysed" lists what could not be checked and why. Pick the last 7, 30
or 90 days. In the app it follows Billing on the Dashboard; in the extension it follows Dashboard in the
panel (GitHub Copilot sessions only). The report changes no setting or file. "Open in Analyze" on
another provider's session switches the selected provider, as the Provider select does. API:
GET /api/savings. See the User Guide's "The savings report". - A subagent marker, block and tile line in the panel. The turns table marks AI Credits that
include subagent runs, the Explanation tab lists the runs with turn totals, and the session tile says
"incl. X from N subagent runs". The MCP App turn detail lists the runs too. - Light and dark colour themes with an Auto/Light/Dark switch in the header; Auto follows the OS setting.
Changed
- Totals now include classic Copilot Chat subagent runs. A classic chat that called a subagent
(runSubagent,searchSubagent) wrote the run to its own debug log, which the analyzer left out.
Each run's AI Credits are now added to the turn that started it, in the turns table, the session
total, the dashboard and the MCP tools. Some totals rise. Some become unknown, with a reason: a
run's log is missing or can't be read, a run can't be matched to a turn, or a session has more than
200 subagent logs. A run whose log no event names is matched by time and labelled so. - Per-model figures, statements and price-change findings credit each run's model. A turn that
ran subagents on another model now counts under each model it used. Statements issued earlier may
show a diff when re-run, because the re-run includes the runs. - Replay compares the main agent. A replayed turn is compared on the main agent's own usage; the
coverage line lists the subagent runs it left out. - Savings action text now writes numbers with thousands separators ("5,400 tokens").
- New look: monospace interface and numbers (Commit Mono), serif explanations (Source Serif 4), square hairline panels; Barlow is no longer bundled.
- Denser layout: smaller spacing and type, 28px table rows, one toolbar row on the Dashboard, compact session and scenario lists that fill the column, 80px KPI tiles with the cost explanation in a footnote, a trend chart that fills its panel, one-line empty charts and a collapsed Learn sources list. The Dashboard is about 40% shorter.
- The server status and version moved into the header.
Fixed
- Opened-turn detail text contrast in VS Code Dark Modern. The opened-turn detail in the MCP App
view now sits on a neutral row background, so its text is readable. - Log records holding the characters U+2028 or U+2029, or a lone carriage return, inside a message are no longer split and dropped: the readers for Claude Code logs, VS Code Copilot debug and event logs, pi sessions, the turn inspector and pi system-prompt sidecars split records at line feeds only, so those sessions' requests, costs and savings figures are complete (found by the savings real-data check; one Claude Code session had lost a third of its requests).
- The MCP server (insights, cost analyses, prices,
estimate_cost,get_statement) now followsAGENT_COST_ANALYZER_NOWlike the web app; it used the real date before, which only differed when the clock is fixed (tests and demos). compare_models_for_sessionand the Replay view no longer label a session that switched models (e.g. Sonnet, then Opus) as measured on its first model alone: the result listsmeasuredModelsin order, flags the first-model entryisMixedSessionModelinstead ofisMeasuredModel, the view says "measured on A, then B" and makes no own-model check;replay_session'sselfCheckgainsmodels, and the self-check line (in the Replay view and the web app) names every model it ran.- On a narrow screen the All sessions audit's tables scroll inside their panels instead of making the page wider.